1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2001 McAfee, Inc. 5 * Copyright (c) 2006,2013 Andre Oppermann, Internet Business Solutions AG 6 * All rights reserved. 7 * 8 * This software was developed for the FreeBSD Project by Jonathan Lemon 9 * and McAfee Research, the Security Research Division of McAfee, Inc. under 10 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the 11 * DARPA CHATS research program. [2001 McAfee, Inc.] 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 #include <sys/cdefs.h> 36 __FBSDID("$FreeBSD$"); 37 38 #include "opt_inet.h" 39 #include "opt_inet6.h" 40 #include "opt_ipsec.h" 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/hash.h> 45 #include <sys/refcount.h> 46 #include <sys/kernel.h> 47 #include <sys/sysctl.h> 48 #include <sys/limits.h> 49 #include <sys/lock.h> 50 #include <sys/mutex.h> 51 #include <sys/malloc.h> 52 #include <sys/mbuf.h> 53 #include <sys/proc.h> /* for proc0 declaration */ 54 #include <sys/random.h> 55 #include <sys/socket.h> 56 #include <sys/socketvar.h> 57 #include <sys/syslog.h> 58 #include <sys/ucred.h> 59 60 #include <sys/md5.h> 61 #include <crypto/siphash/siphash.h> 62 63 #include <vm/uma.h> 64 65 #include <net/if.h> 66 #include <net/if_var.h> 67 #include <net/route.h> 68 #include <net/vnet.h> 69 70 #include <netinet/in.h> 71 #include <netinet/in_kdtrace.h> 72 #include <netinet/in_systm.h> 73 #include <netinet/ip.h> 74 #include <netinet/in_var.h> 75 #include <netinet/in_pcb.h> 76 #include <netinet/ip_var.h> 77 #include <netinet/ip_options.h> 78 #ifdef INET6 79 #include <netinet/ip6.h> 80 #include <netinet/icmp6.h> 81 #include <netinet6/nd6.h> 82 #include <netinet6/ip6_var.h> 83 #include <netinet6/in6_pcb.h> 84 #endif 85 #include <netinet/tcp.h> 86 #include <netinet/tcp_fastopen.h> 87 #include <netinet/tcp_fsm.h> 88 #include <netinet/tcp_seq.h> 89 #include <netinet/tcp_timer.h> 90 #include <netinet/tcp_var.h> 91 #include <netinet/tcp_syncache.h> 92 #include <netinet/tcp_ecn.h> 93 #ifdef TCP_OFFLOAD 94 #include <netinet/toecore.h> 95 #endif 96 #include <netinet/udp.h> 97 98 #include <netipsec/ipsec_support.h> 99 100 #include <machine/in_cksum.h> 101 102 #include <security/mac/mac_framework.h> 103 104 VNET_DEFINE_STATIC(int, tcp_syncookies) = 1; 105 #define V_tcp_syncookies VNET(tcp_syncookies) 106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies, CTLFLAG_VNET | CTLFLAG_RW, 107 &VNET_NAME(tcp_syncookies), 0, 108 "Use TCP SYN cookies if the syncache overflows"); 109 110 VNET_DEFINE_STATIC(int, tcp_syncookiesonly) = 0; 111 #define V_tcp_syncookiesonly VNET(tcp_syncookiesonly) 112 SYSCTL_INT(_net_inet_tcp, OID_AUTO, syncookies_only, CTLFLAG_VNET | CTLFLAG_RW, 113 &VNET_NAME(tcp_syncookiesonly), 0, 114 "Use only TCP SYN cookies"); 115 116 VNET_DEFINE_STATIC(int, functions_inherit_listen_socket_stack) = 1; 117 #define V_functions_inherit_listen_socket_stack \ 118 VNET(functions_inherit_listen_socket_stack) 119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, functions_inherit_listen_socket_stack, 120 CTLFLAG_VNET | CTLFLAG_RW, 121 &VNET_NAME(functions_inherit_listen_socket_stack), 0, 122 "Inherit listen socket's stack"); 123 124 #ifdef TCP_OFFLOAD 125 #define ADDED_BY_TOE(sc) ((sc)->sc_tod != NULL) 126 #endif 127 128 static void syncache_drop(struct syncache *, struct syncache_head *); 129 static void syncache_free(struct syncache *); 130 static void syncache_insert(struct syncache *, struct syncache_head *); 131 static int syncache_respond(struct syncache *, const struct mbuf *, int); 132 static struct socket *syncache_socket(struct syncache *, struct socket *, 133 struct mbuf *m); 134 static void syncache_timeout(struct syncache *sc, struct syncache_head *sch, 135 int docallout); 136 static void syncache_timer(void *); 137 138 static uint32_t syncookie_mac(struct in_conninfo *, tcp_seq, uint8_t, 139 uint8_t *, uintptr_t); 140 static tcp_seq syncookie_generate(struct syncache_head *, struct syncache *); 141 static struct syncache 142 *syncookie_lookup(struct in_conninfo *, struct syncache_head *, 143 struct syncache *, struct tcphdr *, struct tcpopt *, 144 struct socket *, uint16_t); 145 static void syncache_pause(struct in_conninfo *); 146 static void syncache_unpause(void *); 147 static void syncookie_reseed(void *); 148 #ifdef INVARIANTS 149 static int syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 150 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 151 struct socket *lso, uint16_t port); 152 #endif 153 154 /* 155 * Transmit the SYN,ACK fewer times than TCP_MAXRXTSHIFT specifies. 156 * 3 retransmits corresponds to a timeout with default values of 157 * tcp_rexmit_initial * ( 1 + 158 * tcp_backoff[1] + 159 * tcp_backoff[2] + 160 * tcp_backoff[3]) + 3 * tcp_rexmit_slop, 161 * 1000 ms * (1 + 2 + 4 + 8) + 3 * 200 ms = 15600 ms, 162 * the odds are that the user has given up attempting to connect by then. 163 */ 164 #define SYNCACHE_MAXREXMTS 3 165 166 /* Arbitrary values */ 167 #define TCP_SYNCACHE_HASHSIZE 512 168 #define TCP_SYNCACHE_BUCKETLIMIT 30 169 170 VNET_DEFINE_STATIC(struct tcp_syncache, tcp_syncache); 171 #define V_tcp_syncache VNET(tcp_syncache) 172 173 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, syncache, 174 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 175 "TCP SYN cache"); 176 177 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, bucketlimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 178 &VNET_NAME(tcp_syncache.bucket_limit), 0, 179 "Per-bucket hash limit for syncache"); 180 181 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN, 182 &VNET_NAME(tcp_syncache.cache_limit), 0, 183 "Overall entry limit for syncache"); 184 185 SYSCTL_UMA_CUR(_net_inet_tcp_syncache, OID_AUTO, count, CTLFLAG_VNET, 186 &VNET_NAME(tcp_syncache.zone), "Current number of entries in syncache"); 187 188 SYSCTL_UINT(_net_inet_tcp_syncache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN, 189 &VNET_NAME(tcp_syncache.hashsize), 0, 190 "Size of TCP syncache hashtable"); 191 192 SYSCTL_BOOL(_net_inet_tcp_syncache, OID_AUTO, see_other, CTLFLAG_VNET | 193 CTLFLAG_RW, &VNET_NAME(tcp_syncache.see_other), 0, 194 "All syncache(4) entries are visible, ignoring UID/GID, jail(2) " 195 "and mac(4) checks"); 196 197 static int 198 sysctl_net_inet_tcp_syncache_rexmtlimit_check(SYSCTL_HANDLER_ARGS) 199 { 200 int error; 201 u_int new; 202 203 new = V_tcp_syncache.rexmt_limit; 204 error = sysctl_handle_int(oidp, &new, 0, req); 205 if ((error == 0) && (req->newptr != NULL)) { 206 if (new > TCP_MAXRXTSHIFT) 207 error = EINVAL; 208 else 209 V_tcp_syncache.rexmt_limit = new; 210 } 211 return (error); 212 } 213 214 SYSCTL_PROC(_net_inet_tcp_syncache, OID_AUTO, rexmtlimit, 215 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 216 &VNET_NAME(tcp_syncache.rexmt_limit), 0, 217 sysctl_net_inet_tcp_syncache_rexmtlimit_check, "UI", 218 "Limit on SYN/ACK retransmissions"); 219 220 VNET_DEFINE(int, tcp_sc_rst_sock_fail) = 1; 221 SYSCTL_INT(_net_inet_tcp_syncache, OID_AUTO, rst_on_sock_fail, 222 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_sc_rst_sock_fail), 0, 223 "Send reset on socket allocation failure"); 224 225 static MALLOC_DEFINE(M_SYNCACHE, "syncache", "TCP syncache"); 226 227 #define SCH_LOCK(sch) mtx_lock(&(sch)->sch_mtx) 228 #define SCH_UNLOCK(sch) mtx_unlock(&(sch)->sch_mtx) 229 #define SCH_LOCK_ASSERT(sch) mtx_assert(&(sch)->sch_mtx, MA_OWNED) 230 231 /* 232 * Requires the syncache entry to be already removed from the bucket list. 233 */ 234 static void 235 syncache_free(struct syncache *sc) 236 { 237 238 if (sc->sc_ipopts) 239 (void) m_free(sc->sc_ipopts); 240 if (sc->sc_cred) 241 crfree(sc->sc_cred); 242 #ifdef MAC 243 mac_syncache_destroy(&sc->sc_label); 244 #endif 245 246 uma_zfree(V_tcp_syncache.zone, sc); 247 } 248 249 void 250 syncache_init(void) 251 { 252 int i; 253 254 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 255 V_tcp_syncache.bucket_limit = TCP_SYNCACHE_BUCKETLIMIT; 256 V_tcp_syncache.rexmt_limit = SYNCACHE_MAXREXMTS; 257 V_tcp_syncache.hash_secret = arc4random(); 258 259 TUNABLE_INT_FETCH("net.inet.tcp.syncache.hashsize", 260 &V_tcp_syncache.hashsize); 261 TUNABLE_INT_FETCH("net.inet.tcp.syncache.bucketlimit", 262 &V_tcp_syncache.bucket_limit); 263 if (!powerof2(V_tcp_syncache.hashsize) || 264 V_tcp_syncache.hashsize == 0) { 265 printf("WARNING: syncache hash size is not a power of 2.\n"); 266 V_tcp_syncache.hashsize = TCP_SYNCACHE_HASHSIZE; 267 } 268 V_tcp_syncache.hashmask = V_tcp_syncache.hashsize - 1; 269 270 /* Set limits. */ 271 V_tcp_syncache.cache_limit = 272 V_tcp_syncache.hashsize * V_tcp_syncache.bucket_limit; 273 TUNABLE_INT_FETCH("net.inet.tcp.syncache.cachelimit", 274 &V_tcp_syncache.cache_limit); 275 276 /* Allocate the hash table. */ 277 V_tcp_syncache.hashbase = malloc(V_tcp_syncache.hashsize * 278 sizeof(struct syncache_head), M_SYNCACHE, M_WAITOK | M_ZERO); 279 280 #ifdef VIMAGE 281 V_tcp_syncache.vnet = curvnet; 282 #endif 283 284 /* Initialize the hash buckets. */ 285 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 286 TAILQ_INIT(&V_tcp_syncache.hashbase[i].sch_bucket); 287 mtx_init(&V_tcp_syncache.hashbase[i].sch_mtx, "tcp_sc_head", 288 NULL, MTX_DEF); 289 callout_init_mtx(&V_tcp_syncache.hashbase[i].sch_timer, 290 &V_tcp_syncache.hashbase[i].sch_mtx, 0); 291 V_tcp_syncache.hashbase[i].sch_length = 0; 292 V_tcp_syncache.hashbase[i].sch_sc = &V_tcp_syncache; 293 V_tcp_syncache.hashbase[i].sch_last_overflow = 294 -(SYNCOOKIE_LIFETIME + 1); 295 } 296 297 /* Create the syncache entry zone. */ 298 V_tcp_syncache.zone = uma_zcreate("syncache", sizeof(struct syncache), 299 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 300 V_tcp_syncache.cache_limit = uma_zone_set_max(V_tcp_syncache.zone, 301 V_tcp_syncache.cache_limit); 302 303 /* Start the SYN cookie reseeder callout. */ 304 callout_init(&V_tcp_syncache.secret.reseed, 1); 305 arc4rand(V_tcp_syncache.secret.key[0], SYNCOOKIE_SECRET_SIZE, 0); 306 arc4rand(V_tcp_syncache.secret.key[1], SYNCOOKIE_SECRET_SIZE, 0); 307 callout_reset(&V_tcp_syncache.secret.reseed, SYNCOOKIE_LIFETIME * hz, 308 syncookie_reseed, &V_tcp_syncache); 309 310 /* Initialize the pause machinery. */ 311 mtx_init(&V_tcp_syncache.pause_mtx, "tcp_sc_pause", NULL, MTX_DEF); 312 callout_init_mtx(&V_tcp_syncache.pause_co, &V_tcp_syncache.pause_mtx, 313 0); 314 V_tcp_syncache.pause_until = time_uptime - TCP_SYNCACHE_PAUSE_TIME; 315 V_tcp_syncache.pause_backoff = 0; 316 V_tcp_syncache.paused = false; 317 } 318 319 #ifdef VIMAGE 320 void 321 syncache_destroy(void) 322 { 323 struct syncache_head *sch; 324 struct syncache *sc, *nsc; 325 int i; 326 327 /* 328 * Stop the re-seed timer before freeing resources. No need to 329 * possibly schedule it another time. 330 */ 331 callout_drain(&V_tcp_syncache.secret.reseed); 332 333 /* Stop the SYN cache pause callout. */ 334 mtx_lock(&V_tcp_syncache.pause_mtx); 335 if (callout_stop(&V_tcp_syncache.pause_co) == 0) { 336 mtx_unlock(&V_tcp_syncache.pause_mtx); 337 callout_drain(&V_tcp_syncache.pause_co); 338 } else 339 mtx_unlock(&V_tcp_syncache.pause_mtx); 340 341 /* Cleanup hash buckets: stop timers, free entries, destroy locks. */ 342 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 343 sch = &V_tcp_syncache.hashbase[i]; 344 callout_drain(&sch->sch_timer); 345 346 SCH_LOCK(sch); 347 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) 348 syncache_drop(sc, sch); 349 SCH_UNLOCK(sch); 350 KASSERT(TAILQ_EMPTY(&sch->sch_bucket), 351 ("%s: sch->sch_bucket not empty", __func__)); 352 KASSERT(sch->sch_length == 0, ("%s: sch->sch_length %d not 0", 353 __func__, sch->sch_length)); 354 mtx_destroy(&sch->sch_mtx); 355 } 356 357 KASSERT(uma_zone_get_cur(V_tcp_syncache.zone) == 0, 358 ("%s: cache_count not 0", __func__)); 359 360 /* Free the allocated global resources. */ 361 uma_zdestroy(V_tcp_syncache.zone); 362 free(V_tcp_syncache.hashbase, M_SYNCACHE); 363 mtx_destroy(&V_tcp_syncache.pause_mtx); 364 } 365 #endif 366 367 /* 368 * Inserts a syncache entry into the specified bucket row. 369 * Locks and unlocks the syncache_head autonomously. 370 */ 371 static void 372 syncache_insert(struct syncache *sc, struct syncache_head *sch) 373 { 374 struct syncache *sc2; 375 376 SCH_LOCK(sch); 377 378 /* 379 * Make sure that we don't overflow the per-bucket limit. 380 * If the bucket is full, toss the oldest element. 381 */ 382 if (sch->sch_length >= V_tcp_syncache.bucket_limit) { 383 KASSERT(!TAILQ_EMPTY(&sch->sch_bucket), 384 ("sch->sch_length incorrect")); 385 syncache_pause(&sc->sc_inc); 386 sc2 = TAILQ_LAST(&sch->sch_bucket, sch_head); 387 sch->sch_last_overflow = time_uptime; 388 syncache_drop(sc2, sch); 389 } 390 391 /* Put it into the bucket. */ 392 TAILQ_INSERT_HEAD(&sch->sch_bucket, sc, sc_hash); 393 sch->sch_length++; 394 395 #ifdef TCP_OFFLOAD 396 if (ADDED_BY_TOE(sc)) { 397 struct toedev *tod = sc->sc_tod; 398 399 tod->tod_syncache_added(tod, sc->sc_todctx); 400 } 401 #endif 402 403 /* Reinitialize the bucket row's timer. */ 404 if (sch->sch_length == 1) 405 sch->sch_nextc = ticks + INT_MAX; 406 syncache_timeout(sc, sch, 1); 407 408 SCH_UNLOCK(sch); 409 410 TCPSTATES_INC(TCPS_SYN_RECEIVED); 411 TCPSTAT_INC(tcps_sc_added); 412 } 413 414 /* 415 * Remove and free entry from syncache bucket row. 416 * Expects locked syncache head. 417 */ 418 static void 419 syncache_drop(struct syncache *sc, struct syncache_head *sch) 420 { 421 422 SCH_LOCK_ASSERT(sch); 423 424 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 425 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 426 sch->sch_length--; 427 428 #ifdef TCP_OFFLOAD 429 if (ADDED_BY_TOE(sc)) { 430 struct toedev *tod = sc->sc_tod; 431 432 tod->tod_syncache_removed(tod, sc->sc_todctx); 433 } 434 #endif 435 436 syncache_free(sc); 437 } 438 439 /* 440 * Engage/reengage time on bucket row. 441 */ 442 static void 443 syncache_timeout(struct syncache *sc, struct syncache_head *sch, int docallout) 444 { 445 int rexmt; 446 447 if (sc->sc_rxmits == 0) 448 rexmt = tcp_rexmit_initial; 449 else 450 TCPT_RANGESET(rexmt, 451 tcp_rexmit_initial * tcp_backoff[sc->sc_rxmits], 452 tcp_rexmit_min, TCPTV_REXMTMAX); 453 sc->sc_rxttime = ticks + rexmt; 454 sc->sc_rxmits++; 455 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) { 456 sch->sch_nextc = sc->sc_rxttime; 457 if (docallout) 458 callout_reset(&sch->sch_timer, sch->sch_nextc - ticks, 459 syncache_timer, (void *)sch); 460 } 461 } 462 463 /* 464 * Walk the timer queues, looking for SYN,ACKs that need to be retransmitted. 465 * If we have retransmitted an entry the maximum number of times, expire it. 466 * One separate timer for each bucket row. 467 */ 468 static void 469 syncache_timer(void *xsch) 470 { 471 struct syncache_head *sch = (struct syncache_head *)xsch; 472 struct syncache *sc, *nsc; 473 struct epoch_tracker et; 474 int tick = ticks; 475 char *s; 476 bool paused; 477 478 CURVNET_SET(sch->sch_sc->vnet); 479 480 /* NB: syncache_head has already been locked by the callout. */ 481 SCH_LOCK_ASSERT(sch); 482 483 /* 484 * In the following cycle we may remove some entries and/or 485 * advance some timeouts, so re-initialize the bucket timer. 486 */ 487 sch->sch_nextc = tick + INT_MAX; 488 489 /* 490 * If we have paused processing, unconditionally remove 491 * all syncache entries. 492 */ 493 mtx_lock(&V_tcp_syncache.pause_mtx); 494 paused = V_tcp_syncache.paused; 495 mtx_unlock(&V_tcp_syncache.pause_mtx); 496 497 TAILQ_FOREACH_SAFE(sc, &sch->sch_bucket, sc_hash, nsc) { 498 if (paused) { 499 syncache_drop(sc, sch); 500 continue; 501 } 502 /* 503 * We do not check if the listen socket still exists 504 * and accept the case where the listen socket may be 505 * gone by the time we resend the SYN/ACK. We do 506 * not expect this to happens often. If it does, 507 * then the RST will be sent by the time the remote 508 * host does the SYN/ACK->ACK. 509 */ 510 if (TSTMP_GT(sc->sc_rxttime, tick)) { 511 if (TSTMP_LT(sc->sc_rxttime, sch->sch_nextc)) 512 sch->sch_nextc = sc->sc_rxttime; 513 continue; 514 } 515 if (sc->sc_rxmits > V_tcp_ecn_maxretries) { 516 sc->sc_flags &= ~SCF_ECN_MASK; 517 } 518 if (sc->sc_rxmits > V_tcp_syncache.rexmt_limit) { 519 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 520 log(LOG_DEBUG, "%s; %s: Retransmits exhausted, " 521 "giving up and removing syncache entry\n", 522 s, __func__); 523 free(s, M_TCPLOG); 524 } 525 syncache_drop(sc, sch); 526 TCPSTAT_INC(tcps_sc_stale); 527 continue; 528 } 529 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 530 log(LOG_DEBUG, "%s; %s: Response timeout, " 531 "retransmitting (%u) SYN|ACK\n", 532 s, __func__, sc->sc_rxmits); 533 free(s, M_TCPLOG); 534 } 535 536 NET_EPOCH_ENTER(et); 537 syncache_respond(sc, NULL, TH_SYN|TH_ACK); 538 NET_EPOCH_EXIT(et); 539 TCPSTAT_INC(tcps_sc_retransmitted); 540 syncache_timeout(sc, sch, 0); 541 } 542 if (!TAILQ_EMPTY(&(sch)->sch_bucket)) 543 callout_reset(&(sch)->sch_timer, (sch)->sch_nextc - tick, 544 syncache_timer, (void *)(sch)); 545 CURVNET_RESTORE(); 546 } 547 548 /* 549 * Returns true if the system is only using cookies at the moment. 550 * This could be due to a sysadmin decision to only use cookies, or it 551 * could be due to the system detecting an attack. 552 */ 553 static inline bool 554 syncache_cookiesonly(void) 555 { 556 557 return (V_tcp_syncookies && (V_tcp_syncache.paused || 558 V_tcp_syncookiesonly)); 559 } 560 561 /* 562 * Find the hash bucket for the given connection. 563 */ 564 static struct syncache_head * 565 syncache_hashbucket(struct in_conninfo *inc) 566 { 567 uint32_t hash; 568 569 /* 570 * The hash is built on foreign port + local port + foreign address. 571 * We rely on the fact that struct in_conninfo starts with 16 bits 572 * of foreign port, then 16 bits of local port then followed by 128 573 * bits of foreign address. In case of IPv4 address, the first 3 574 * 32-bit words of the address always are zeroes. 575 */ 576 hash = jenkins_hash32((uint32_t *)&inc->inc_ie, 5, 577 V_tcp_syncache.hash_secret) & V_tcp_syncache.hashmask; 578 579 return (&V_tcp_syncache.hashbase[hash]); 580 } 581 582 /* 583 * Find an entry in the syncache. 584 * Returns always with locked syncache_head plus a matching entry or NULL. 585 */ 586 static struct syncache * 587 syncache_lookup(struct in_conninfo *inc, struct syncache_head **schp) 588 { 589 struct syncache *sc; 590 struct syncache_head *sch; 591 592 *schp = sch = syncache_hashbucket(inc); 593 SCH_LOCK(sch); 594 595 /* Circle through bucket row to find matching entry. */ 596 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) 597 if (bcmp(&inc->inc_ie, &sc->sc_inc.inc_ie, 598 sizeof(struct in_endpoints)) == 0) 599 break; 600 601 return (sc); /* Always returns with locked sch. */ 602 } 603 604 /* 605 * This function is called when we get a RST for a 606 * non-existent connection, so that we can see if the 607 * connection is in the syn cache. If it is, zap it. 608 * If required send a challenge ACK. 609 */ 610 void 611 syncache_chkrst(struct in_conninfo *inc, struct tcphdr *th, struct mbuf *m, 612 uint16_t port) 613 { 614 struct syncache *sc; 615 struct syncache_head *sch; 616 char *s = NULL; 617 618 if (syncache_cookiesonly()) 619 return; 620 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 621 SCH_LOCK_ASSERT(sch); 622 623 /* 624 * Any RST to our SYN|ACK must not carry ACK, SYN or FIN flags. 625 * See RFC 793 page 65, section SEGMENT ARRIVES. 626 */ 627 if (tcp_get_flags(th) & (TH_ACK|TH_SYN|TH_FIN)) { 628 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 629 log(LOG_DEBUG, "%s; %s: Spurious RST with ACK, SYN or " 630 "FIN flag set, segment ignored\n", s, __func__); 631 TCPSTAT_INC(tcps_badrst); 632 goto done; 633 } 634 635 /* 636 * No corresponding connection was found in syncache. 637 * If syncookies are enabled and possibly exclusively 638 * used, or we are under memory pressure, a valid RST 639 * may not find a syncache entry. In that case we're 640 * done and no SYN|ACK retransmissions will happen. 641 * Otherwise the RST was misdirected or spoofed. 642 */ 643 if (sc == NULL) { 644 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 645 log(LOG_DEBUG, "%s; %s: Spurious RST without matching " 646 "syncache entry (possibly syncookie only), " 647 "segment ignored\n", s, __func__); 648 TCPSTAT_INC(tcps_badrst); 649 goto done; 650 } 651 652 /* The remote UDP encaps port does not match. */ 653 if (sc->sc_port != port) { 654 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 655 log(LOG_DEBUG, "%s; %s: Spurious RST with matching " 656 "syncache entry but non-matching UDP encaps port, " 657 "segment ignored\n", s, __func__); 658 TCPSTAT_INC(tcps_badrst); 659 goto done; 660 } 661 662 /* 663 * If the RST bit is set, check the sequence number to see 664 * if this is a valid reset segment. 665 * 666 * RFC 793 page 37: 667 * In all states except SYN-SENT, all reset (RST) segments 668 * are validated by checking their SEQ-fields. A reset is 669 * valid if its sequence number is in the window. 670 * 671 * RFC 793 page 69: 672 * There are four cases for the acceptability test for an incoming 673 * segment: 674 * 675 * Segment Receive Test 676 * Length Window 677 * ------- ------- ------------------------------------------- 678 * 0 0 SEG.SEQ = RCV.NXT 679 * 0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 680 * >0 0 not acceptable 681 * >0 >0 RCV.NXT =< SEG.SEQ < RCV.NXT+RCV.WND 682 * or RCV.NXT =< SEG.SEQ+SEG.LEN-1 < RCV.NXT+RCV.WND 683 * 684 * Note that when receiving a SYN segment in the LISTEN state, 685 * IRS is set to SEG.SEQ and RCV.NXT is set to SEG.SEQ+1, as 686 * described in RFC 793, page 66. 687 */ 688 if ((SEQ_GEQ(th->th_seq, sc->sc_irs + 1) && 689 SEQ_LT(th->th_seq, sc->sc_irs + 1 + sc->sc_wnd)) || 690 (sc->sc_wnd == 0 && th->th_seq == sc->sc_irs + 1)) { 691 if (V_tcp_insecure_rst || 692 th->th_seq == sc->sc_irs + 1) { 693 syncache_drop(sc, sch); 694 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 695 log(LOG_DEBUG, 696 "%s; %s: Our SYN|ACK was rejected, " 697 "connection attempt aborted by remote " 698 "endpoint\n", 699 s, __func__); 700 TCPSTAT_INC(tcps_sc_reset); 701 } else { 702 TCPSTAT_INC(tcps_badrst); 703 /* Send challenge ACK. */ 704 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 705 log(LOG_DEBUG, "%s; %s: RST with invalid " 706 " SEQ %u != NXT %u (+WND %u), " 707 "sending challenge ACK\n", 708 s, __func__, 709 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 710 syncache_respond(sc, m, TH_ACK); 711 } 712 } else { 713 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 714 log(LOG_DEBUG, "%s; %s: RST with invalid SEQ %u != " 715 "NXT %u (+WND %u), segment ignored\n", 716 s, __func__, 717 th->th_seq, sc->sc_irs + 1, sc->sc_wnd); 718 TCPSTAT_INC(tcps_badrst); 719 } 720 721 done: 722 if (s != NULL) 723 free(s, M_TCPLOG); 724 SCH_UNLOCK(sch); 725 } 726 727 void 728 syncache_badack(struct in_conninfo *inc, uint16_t port) 729 { 730 struct syncache *sc; 731 struct syncache_head *sch; 732 733 if (syncache_cookiesonly()) 734 return; 735 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 736 SCH_LOCK_ASSERT(sch); 737 if ((sc != NULL) && (sc->sc_port == port)) { 738 syncache_drop(sc, sch); 739 TCPSTAT_INC(tcps_sc_badack); 740 } 741 SCH_UNLOCK(sch); 742 } 743 744 void 745 syncache_unreach(struct in_conninfo *inc, tcp_seq th_seq, uint16_t port) 746 { 747 struct syncache *sc; 748 struct syncache_head *sch; 749 750 if (syncache_cookiesonly()) 751 return; 752 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 753 SCH_LOCK_ASSERT(sch); 754 if (sc == NULL) 755 goto done; 756 757 /* If the port != sc_port, then it's a bogus ICMP msg */ 758 if (port != sc->sc_port) 759 goto done; 760 761 /* If the sequence number != sc_iss, then it's a bogus ICMP msg */ 762 if (ntohl(th_seq) != sc->sc_iss) 763 goto done; 764 765 /* 766 * If we've rertransmitted 3 times and this is our second error, 767 * we remove the entry. Otherwise, we allow it to continue on. 768 * This prevents us from incorrectly nuking an entry during a 769 * spurious network outage. 770 * 771 * See tcp_notify(). 772 */ 773 if ((sc->sc_flags & SCF_UNREACH) == 0 || sc->sc_rxmits < 3 + 1) { 774 sc->sc_flags |= SCF_UNREACH; 775 goto done; 776 } 777 syncache_drop(sc, sch); 778 TCPSTAT_INC(tcps_sc_unreach); 779 done: 780 SCH_UNLOCK(sch); 781 } 782 783 /* 784 * Build a new TCP socket structure from a syncache entry. 785 * 786 * On success return the newly created socket with its underlying inp locked. 787 */ 788 static struct socket * 789 syncache_socket(struct syncache *sc, struct socket *lso, struct mbuf *m) 790 { 791 struct tcp_function_block *blk; 792 struct inpcb *inp = NULL; 793 struct socket *so; 794 struct tcpcb *tp; 795 int error; 796 char *s; 797 798 NET_EPOCH_ASSERT(); 799 800 /* 801 * Ok, create the full blown connection, and set things up 802 * as they would have been set up if we had created the 803 * connection when the SYN arrived. 804 */ 805 if ((so = solisten_clone(lso)) == NULL) 806 goto allocfail; 807 #ifdef MAC 808 mac_socketpeer_set_from_mbuf(m, so); 809 #endif 810 error = in_pcballoc(so, &V_tcbinfo); 811 if (error) { 812 sodealloc(so); 813 goto allocfail; 814 } 815 inp = sotoinpcb(so); 816 if ((tp = tcp_newtcpcb(inp)) == NULL) { 817 in_pcbdetach(inp); 818 in_pcbfree(inp); 819 sodealloc(so); 820 goto allocfail; 821 } 822 inp->inp_inc.inc_flags = sc->sc_inc.inc_flags; 823 #ifdef INET6 824 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 825 inp->inp_vflag &= ~INP_IPV4; 826 inp->inp_vflag |= INP_IPV6; 827 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 828 } else { 829 inp->inp_vflag &= ~INP_IPV6; 830 inp->inp_vflag |= INP_IPV4; 831 #endif 832 inp->inp_ip_ttl = sc->sc_ip_ttl; 833 inp->inp_ip_tos = sc->sc_ip_tos; 834 inp->inp_laddr = sc->sc_inc.inc_laddr; 835 #ifdef INET6 836 } 837 #endif 838 839 /* 840 * If there's an mbuf and it has a flowid, then let's initialise the 841 * inp with that particular flowid. 842 */ 843 if (m != NULL && M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 844 inp->inp_flowid = m->m_pkthdr.flowid; 845 inp->inp_flowtype = M_HASHTYPE_GET(m); 846 #ifdef NUMA 847 inp->inp_numa_domain = m->m_pkthdr.numa_domain; 848 #endif 849 } 850 851 inp->inp_lport = sc->sc_inc.inc_lport; 852 #ifdef INET6 853 if (inp->inp_vflag & INP_IPV6PROTO) { 854 struct inpcb *oinp = sotoinpcb(lso); 855 856 /* 857 * Inherit socket options from the listening socket. 858 * Note that in6p_inputopts are not (and should not be) 859 * copied, since it stores previously received options and is 860 * used to detect if each new option is different than the 861 * previous one and hence should be passed to a user. 862 * If we copied in6p_inputopts, a user would not be able to 863 * receive options just after calling the accept system call. 864 */ 865 inp->inp_flags |= oinp->inp_flags & INP_CONTROLOPTS; 866 if (oinp->in6p_outputopts) 867 inp->in6p_outputopts = 868 ip6_copypktopts(oinp->in6p_outputopts, M_NOWAIT); 869 inp->in6p_hops = oinp->in6p_hops; 870 } 871 872 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 873 struct in6_addr laddr6; 874 struct sockaddr_in6 sin6; 875 876 sin6.sin6_family = AF_INET6; 877 sin6.sin6_len = sizeof(sin6); 878 sin6.sin6_addr = sc->sc_inc.inc6_faddr; 879 sin6.sin6_port = sc->sc_inc.inc_fport; 880 sin6.sin6_flowinfo = sin6.sin6_scope_id = 0; 881 laddr6 = inp->in6p_laddr; 882 if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) 883 inp->in6p_laddr = sc->sc_inc.inc6_laddr; 884 INP_HASH_WLOCK(&V_tcbinfo); 885 error = in6_pcbconnect(inp, &sin6, thread0.td_ucred, false); 886 INP_HASH_WUNLOCK(&V_tcbinfo); 887 if (error != 0) { 888 inp->in6p_laddr = laddr6; 889 goto abort; 890 } 891 /* Override flowlabel from in6_pcbconnect. */ 892 inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; 893 inp->inp_flow |= sc->sc_flowlabel; 894 } 895 #endif /* INET6 */ 896 #if defined(INET) && defined(INET6) 897 else 898 #endif 899 #ifdef INET 900 { 901 struct in_addr laddr; 902 struct sockaddr_in sin; 903 904 inp->inp_options = (m) ? ip_srcroute(m) : NULL; 905 906 if (inp->inp_options == NULL) { 907 inp->inp_options = sc->sc_ipopts; 908 sc->sc_ipopts = NULL; 909 } 910 911 sin.sin_family = AF_INET; 912 sin.sin_len = sizeof(sin); 913 sin.sin_addr = sc->sc_inc.inc_faddr; 914 sin.sin_port = sc->sc_inc.inc_fport; 915 bzero((caddr_t)sin.sin_zero, sizeof(sin.sin_zero)); 916 laddr = inp->inp_laddr; 917 if (inp->inp_laddr.s_addr == INADDR_ANY) 918 inp->inp_laddr = sc->sc_inc.inc_laddr; 919 INP_HASH_WLOCK(&V_tcbinfo); 920 error = in_pcbconnect(inp, &sin, thread0.td_ucred, false); 921 INP_HASH_WUNLOCK(&V_tcbinfo); 922 if (error != 0) { 923 inp->inp_laddr = laddr; 924 goto abort; 925 } 926 } 927 #endif /* INET */ 928 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 929 /* Copy old policy into new socket's. */ 930 if (ipsec_copy_pcbpolicy(sotoinpcb(lso), inp) != 0) 931 printf("syncache_socket: could not copy policy\n"); 932 #endif 933 tp->t_state = TCPS_SYN_RECEIVED; 934 tp->iss = sc->sc_iss; 935 tp->irs = sc->sc_irs; 936 tp->t_port = sc->sc_port; 937 tcp_rcvseqinit(tp); 938 tcp_sendseqinit(tp); 939 blk = sototcpcb(lso)->t_fb; 940 if (V_functions_inherit_listen_socket_stack && blk != tp->t_fb) { 941 /* 942 * Our parents t_fb was not the default, 943 * we need to release our ref on tp->t_fb and 944 * pickup one on the new entry. 945 */ 946 struct tcp_function_block *rblk; 947 948 rblk = find_and_ref_tcp_fb(blk); 949 KASSERT(rblk != NULL, 950 ("cannot find blk %p out of syncache?", blk)); 951 if (tp->t_fb->tfb_tcp_fb_fini) 952 (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); 953 refcount_release(&tp->t_fb->tfb_refcnt); 954 tp->t_fb = rblk; 955 /* 956 * XXXrrs this is quite dangerous, it is possible 957 * for the new function to fail to init. We also 958 * are not asking if the handoff_is_ok though at 959 * the very start thats probalbly ok. 960 */ 961 if (tp->t_fb->tfb_tcp_fb_init) { 962 (*tp->t_fb->tfb_tcp_fb_init)(tp); 963 } 964 } 965 tp->snd_wl1 = sc->sc_irs; 966 tp->snd_max = tp->iss + 1; 967 tp->snd_nxt = tp->iss + 1; 968 tp->rcv_up = sc->sc_irs + 1; 969 tp->rcv_wnd = sc->sc_wnd; 970 tp->rcv_adv += tp->rcv_wnd; 971 tp->last_ack_sent = tp->rcv_nxt; 972 973 tp->t_flags = sototcpcb(lso)->t_flags & (TF_NOPUSH|TF_NODELAY); 974 if (sc->sc_flags & SCF_NOOPT) 975 tp->t_flags |= TF_NOOPT; 976 else { 977 if (sc->sc_flags & SCF_WINSCALE) { 978 tp->t_flags |= TF_REQ_SCALE|TF_RCVD_SCALE; 979 tp->snd_scale = sc->sc_requested_s_scale; 980 tp->request_r_scale = sc->sc_requested_r_scale; 981 } 982 if (sc->sc_flags & SCF_TIMESTAMP) { 983 tp->t_flags |= TF_REQ_TSTMP|TF_RCVD_TSTMP; 984 tp->ts_recent = sc->sc_tsreflect; 985 tp->ts_recent_age = tcp_ts_getticks(); 986 tp->ts_offset = sc->sc_tsoff; 987 } 988 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 989 if (sc->sc_flags & SCF_SIGNATURE) 990 tp->t_flags |= TF_SIGNATURE; 991 #endif 992 if (sc->sc_flags & SCF_SACK) 993 tp->t_flags |= TF_SACK_PERMIT; 994 } 995 996 tcp_ecn_syncache_socket(tp, sc); 997 998 /* 999 * Set up MSS and get cached values from tcp_hostcache. 1000 * This might overwrite some of the defaults we just set. 1001 */ 1002 tcp_mss(tp, sc->sc_peer_mss); 1003 1004 /* 1005 * If the SYN,ACK was retransmitted, indicate that CWND to be 1006 * limited to one segment in cc_conn_init(). 1007 * NB: sc_rxmits counts all SYN,ACK transmits, not just retransmits. 1008 */ 1009 if (sc->sc_rxmits > 1) 1010 tp->snd_cwnd = 1; 1011 1012 #ifdef TCP_OFFLOAD 1013 /* 1014 * Allow a TOE driver to install its hooks. Note that we hold the 1015 * pcbinfo lock too and that prevents tcp_usr_accept from accepting a 1016 * new connection before the TOE driver has done its thing. 1017 */ 1018 if (ADDED_BY_TOE(sc)) { 1019 struct toedev *tod = sc->sc_tod; 1020 1021 tod->tod_offload_socket(tod, sc->sc_todctx, so); 1022 } 1023 #endif 1024 /* 1025 * Copy and activate timers. 1026 */ 1027 tp->t_maxunacktime = sototcpcb(lso)->t_maxunacktime; 1028 tp->t_keepinit = sototcpcb(lso)->t_keepinit; 1029 tp->t_keepidle = sototcpcb(lso)->t_keepidle; 1030 tp->t_keepintvl = sototcpcb(lso)->t_keepintvl; 1031 tp->t_keepcnt = sototcpcb(lso)->t_keepcnt; 1032 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); 1033 1034 TCPSTAT_INC(tcps_accepts); 1035 TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, TCPS_LISTEN); 1036 1037 if (!solisten_enqueue(so, SS_ISCONNECTED)) 1038 tp->t_flags |= TF_SONOTCONN; 1039 1040 return (so); 1041 1042 allocfail: 1043 /* 1044 * Drop the connection; we will either send a RST or have the peer 1045 * retransmit its SYN again after its RTO and try again. 1046 */ 1047 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 1048 log(LOG_DEBUG, "%s; %s: Socket create failed " 1049 "due to limits or memory shortage\n", 1050 s, __func__); 1051 free(s, M_TCPLOG); 1052 } 1053 TCPSTAT_INC(tcps_listendrop); 1054 return (NULL); 1055 1056 abort: 1057 in_pcbdetach(inp); 1058 in_pcbfree(inp); 1059 sodealloc(so); 1060 if ((s = tcp_log_addrs(&sc->sc_inc, NULL, NULL, NULL))) { 1061 log(LOG_DEBUG, "%s; %s: in%s_pcbconnect failed with error %i\n", 1062 s, __func__, (sc->sc_inc.inc_flags & INC_ISIPV6) ? "6" : "", 1063 error); 1064 free(s, M_TCPLOG); 1065 } 1066 TCPSTAT_INC(tcps_listendrop); 1067 return (NULL); 1068 } 1069 1070 /* 1071 * This function gets called when we receive an ACK for a 1072 * socket in the LISTEN state. We look up the connection 1073 * in the syncache, and if its there, we pull it out of 1074 * the cache and turn it into a full-blown connection in 1075 * the SYN-RECEIVED state. 1076 * 1077 * On syncache_socket() success the newly created socket 1078 * has its underlying inp locked. 1079 */ 1080 int 1081 syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1082 struct socket **lsop, struct mbuf *m, uint16_t port) 1083 { 1084 struct syncache *sc; 1085 struct syncache_head *sch; 1086 struct syncache scs; 1087 char *s; 1088 bool locked; 1089 1090 NET_EPOCH_ASSERT(); 1091 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK, 1092 ("%s: can handle only ACK", __func__)); 1093 1094 if (syncache_cookiesonly()) { 1095 sc = NULL; 1096 sch = syncache_hashbucket(inc); 1097 locked = false; 1098 } else { 1099 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1100 locked = true; 1101 SCH_LOCK_ASSERT(sch); 1102 } 1103 1104 #ifdef INVARIANTS 1105 /* 1106 * Test code for syncookies comparing the syncache stored 1107 * values with the reconstructed values from the cookie. 1108 */ 1109 if (sc != NULL) 1110 syncookie_cmp(inc, sch, sc, th, to, *lsop, port); 1111 #endif 1112 1113 if (sc == NULL) { 1114 /* 1115 * There is no syncache entry, so see if this ACK is 1116 * a returning syncookie. To do this, first: 1117 * A. Check if syncookies are used in case of syncache 1118 * overflows 1119 * B. See if this socket has had a syncache entry dropped in 1120 * the recent past. We don't want to accept a bogus 1121 * syncookie if we've never received a SYN or accept it 1122 * twice. 1123 * C. check that the syncookie is valid. If it is, then 1124 * cobble up a fake syncache entry, and return. 1125 */ 1126 if (locked && !V_tcp_syncookies) { 1127 SCH_UNLOCK(sch); 1128 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1129 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1130 "segment rejected (syncookies disabled)\n", 1131 s, __func__); 1132 goto failed; 1133 } 1134 if (locked && !V_tcp_syncookiesonly && 1135 sch->sch_last_overflow < time_uptime - SYNCOOKIE_LIFETIME) { 1136 SCH_UNLOCK(sch); 1137 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1138 log(LOG_DEBUG, "%s; %s: Spurious ACK, " 1139 "segment rejected (no syncache entry)\n", 1140 s, __func__); 1141 goto failed; 1142 } 1143 bzero(&scs, sizeof(scs)); 1144 sc = syncookie_lookup(inc, sch, &scs, th, to, *lsop, port); 1145 if (locked) 1146 SCH_UNLOCK(sch); 1147 if (sc == NULL) { 1148 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1149 log(LOG_DEBUG, "%s; %s: Segment failed " 1150 "SYNCOOKIE authentication, segment rejected " 1151 "(probably spoofed)\n", s, __func__); 1152 goto failed; 1153 } 1154 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1155 /* If received ACK has MD5 signature, check it. */ 1156 if ((to->to_flags & TOF_SIGNATURE) != 0 && 1157 (!TCPMD5_ENABLED() || 1158 TCPMD5_INPUT(m, th, to->to_signature) != 0)) { 1159 /* Drop the ACK. */ 1160 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1161 log(LOG_DEBUG, "%s; %s: Segment rejected, " 1162 "MD5 signature doesn't match.\n", 1163 s, __func__); 1164 free(s, M_TCPLOG); 1165 } 1166 TCPSTAT_INC(tcps_sig_err_sigopt); 1167 return (-1); /* Do not send RST */ 1168 } 1169 #endif /* TCP_SIGNATURE */ 1170 TCPSTATES_INC(TCPS_SYN_RECEIVED); 1171 } else { 1172 if (sc->sc_port != port) { 1173 SCH_UNLOCK(sch); 1174 return (0); 1175 } 1176 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1177 /* 1178 * If listening socket requested TCP digests, check that 1179 * received ACK has signature and it is correct. 1180 * If not, drop the ACK and leave sc entry in th cache, 1181 * because SYN was received with correct signature. 1182 */ 1183 if (sc->sc_flags & SCF_SIGNATURE) { 1184 if ((to->to_flags & TOF_SIGNATURE) == 0) { 1185 /* No signature */ 1186 TCPSTAT_INC(tcps_sig_err_nosigopt); 1187 SCH_UNLOCK(sch); 1188 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1189 log(LOG_DEBUG, "%s; %s: Segment " 1190 "rejected, MD5 signature wasn't " 1191 "provided.\n", s, __func__); 1192 free(s, M_TCPLOG); 1193 } 1194 return (-1); /* Do not send RST */ 1195 } 1196 if (!TCPMD5_ENABLED() || 1197 TCPMD5_INPUT(m, th, to->to_signature) != 0) { 1198 /* Doesn't match or no SA */ 1199 SCH_UNLOCK(sch); 1200 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1201 log(LOG_DEBUG, "%s; %s: Segment " 1202 "rejected, MD5 signature doesn't " 1203 "match.\n", s, __func__); 1204 free(s, M_TCPLOG); 1205 } 1206 return (-1); /* Do not send RST */ 1207 } 1208 } 1209 #endif /* TCP_SIGNATURE */ 1210 1211 /* 1212 * RFC 7323 PAWS: If we have a timestamp on this segment and 1213 * it's less than ts_recent, drop it. 1214 * XXXMT: RFC 7323 also requires to send an ACK. 1215 * In tcp_input.c this is only done for TCP segments 1216 * with user data, so be consistent here and just drop 1217 * the segment. 1218 */ 1219 if (sc->sc_flags & SCF_TIMESTAMP && to->to_flags & TOF_TS && 1220 TSTMP_LT(to->to_tsval, sc->sc_tsreflect)) { 1221 SCH_UNLOCK(sch); 1222 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1223 log(LOG_DEBUG, 1224 "%s; %s: SEG.TSval %u < TS.Recent %u, " 1225 "segment dropped\n", s, __func__, 1226 to->to_tsval, sc->sc_tsreflect); 1227 free(s, M_TCPLOG); 1228 } 1229 return (-1); /* Do not send RST */ 1230 } 1231 1232 /* 1233 * If timestamps were not negotiated during SYN/ACK and a 1234 * segment with a timestamp is received, ignore the 1235 * timestamp and process the packet normally. 1236 * See section 3.2 of RFC 7323. 1237 */ 1238 if (!(sc->sc_flags & SCF_TIMESTAMP) && 1239 (to->to_flags & TOF_TS)) { 1240 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1241 log(LOG_DEBUG, "%s; %s: Timestamp not " 1242 "expected, segment processed normally\n", 1243 s, __func__); 1244 free(s, M_TCPLOG); 1245 s = NULL; 1246 } 1247 } 1248 1249 /* 1250 * If timestamps were negotiated during SYN/ACK and a 1251 * segment without a timestamp is received, silently drop 1252 * the segment, unless the missing timestamps are tolerated. 1253 * See section 3.2 of RFC 7323. 1254 */ 1255 if ((sc->sc_flags & SCF_TIMESTAMP) && 1256 !(to->to_flags & TOF_TS)) { 1257 if (V_tcp_tolerate_missing_ts) { 1258 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1259 log(LOG_DEBUG, 1260 "%s; %s: Timestamp missing, " 1261 "segment processed normally\n", 1262 s, __func__); 1263 free(s, M_TCPLOG); 1264 } 1265 } else { 1266 SCH_UNLOCK(sch); 1267 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { 1268 log(LOG_DEBUG, 1269 "%s; %s: Timestamp missing, " 1270 "segment silently dropped\n", 1271 s, __func__); 1272 free(s, M_TCPLOG); 1273 } 1274 return (-1); /* Do not send RST */ 1275 } 1276 } 1277 TAILQ_REMOVE(&sch->sch_bucket, sc, sc_hash); 1278 sch->sch_length--; 1279 #ifdef TCP_OFFLOAD 1280 if (ADDED_BY_TOE(sc)) { 1281 struct toedev *tod = sc->sc_tod; 1282 1283 tod->tod_syncache_removed(tod, sc->sc_todctx); 1284 } 1285 #endif 1286 SCH_UNLOCK(sch); 1287 } 1288 1289 /* 1290 * Segment validation: 1291 * ACK must match our initial sequence number + 1 (the SYN|ACK). 1292 */ 1293 if (th->th_ack != sc->sc_iss + 1) { 1294 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1295 log(LOG_DEBUG, "%s; %s: ACK %u != ISS+1 %u, segment " 1296 "rejected\n", s, __func__, th->th_ack, sc->sc_iss); 1297 goto failed; 1298 } 1299 1300 /* 1301 * The SEQ must fall in the window starting at the received 1302 * initial receive sequence number + 1 (the SYN). 1303 */ 1304 if (SEQ_LEQ(th->th_seq, sc->sc_irs) || 1305 SEQ_GT(th->th_seq, sc->sc_irs + sc->sc_wnd)) { 1306 if ((s = tcp_log_addrs(inc, th, NULL, NULL))) 1307 log(LOG_DEBUG, "%s; %s: SEQ %u != IRS+1 %u, segment " 1308 "rejected\n", s, __func__, th->th_seq, sc->sc_irs); 1309 goto failed; 1310 } 1311 1312 *lsop = syncache_socket(sc, *lsop, m); 1313 1314 if (__predict_false(*lsop == NULL)) { 1315 TCPSTAT_INC(tcps_sc_aborted); 1316 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1317 } else 1318 TCPSTAT_INC(tcps_sc_completed); 1319 1320 /* how do we find the inp for the new socket? */ 1321 if (sc != &scs) 1322 syncache_free(sc); 1323 return (1); 1324 failed: 1325 if (sc != NULL) { 1326 TCPSTATES_DEC(TCPS_SYN_RECEIVED); 1327 if (sc != &scs) 1328 syncache_free(sc); 1329 } 1330 if (s != NULL) 1331 free(s, M_TCPLOG); 1332 *lsop = NULL; 1333 return (0); 1334 } 1335 1336 static struct socket * 1337 syncache_tfo_expand(struct syncache *sc, struct socket *lso, struct mbuf *m, 1338 uint64_t response_cookie) 1339 { 1340 struct inpcb *inp; 1341 struct tcpcb *tp; 1342 unsigned int *pending_counter; 1343 struct socket *so; 1344 1345 NET_EPOCH_ASSERT(); 1346 1347 pending_counter = intotcpcb(sotoinpcb(lso))->t_tfo_pending; 1348 so = syncache_socket(sc, lso, m); 1349 if (so == NULL) { 1350 TCPSTAT_INC(tcps_sc_aborted); 1351 atomic_subtract_int(pending_counter, 1); 1352 } else { 1353 soisconnected(so); 1354 inp = sotoinpcb(so); 1355 tp = intotcpcb(inp); 1356 tp->t_flags |= TF_FASTOPEN; 1357 tp->t_tfo_cookie.server = response_cookie; 1358 tp->snd_max = tp->iss; 1359 tp->snd_nxt = tp->iss; 1360 tp->t_tfo_pending = pending_counter; 1361 TCPSTATES_INC(TCPS_SYN_RECEIVED); 1362 TCPSTAT_INC(tcps_sc_completed); 1363 } 1364 1365 return (so); 1366 } 1367 1368 /* 1369 * Given a LISTEN socket and an inbound SYN request, add 1370 * this to the syn cache, and send back a segment: 1371 * <SEQ=ISS><ACK=RCV_NXT><CTL=SYN,ACK> 1372 * to the source. 1373 * 1374 * IMPORTANT NOTE: We do _NOT_ ACK data that might accompany the SYN. 1375 * Doing so would require that we hold onto the data and deliver it 1376 * to the application. However, if we are the target of a SYN-flood 1377 * DoS attack, an attacker could send data which would eventually 1378 * consume all available buffer space if it were ACKed. By not ACKing 1379 * the data, we avoid this DoS scenario. 1380 * 1381 * The exception to the above is when a SYN with a valid TCP Fast Open (TFO) 1382 * cookie is processed and a new socket is created. In this case, any data 1383 * accompanying the SYN will be queued to the socket by tcp_input() and will 1384 * be ACKed either when the application sends response data or the delayed 1385 * ACK timer expires, whichever comes first. 1386 */ 1387 struct socket * 1388 syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, 1389 struct inpcb *inp, struct socket *so, struct mbuf *m, void *tod, 1390 void *todctx, uint8_t iptos, uint16_t port) 1391 { 1392 struct tcpcb *tp; 1393 struct socket *rv = NULL; 1394 struct syncache *sc = NULL; 1395 struct syncache_head *sch; 1396 struct mbuf *ipopts = NULL; 1397 u_int ltflags; 1398 int win, ip_ttl, ip_tos; 1399 char *s; 1400 #ifdef INET6 1401 int autoflowlabel = 0; 1402 #endif 1403 #ifdef MAC 1404 struct label *maclabel; 1405 #endif 1406 struct syncache scs; 1407 struct ucred *cred; 1408 uint64_t tfo_response_cookie; 1409 unsigned int *tfo_pending = NULL; 1410 int tfo_cookie_valid = 0; 1411 int tfo_response_cookie_valid = 0; 1412 bool locked; 1413 1414 INP_RLOCK_ASSERT(inp); /* listen socket */ 1415 KASSERT((tcp_get_flags(th) & (TH_RST|TH_ACK|TH_SYN)) == TH_SYN, 1416 ("%s: unexpected tcp flags", __func__)); 1417 1418 /* 1419 * Combine all so/tp operations very early to drop the INP lock as 1420 * soon as possible. 1421 */ 1422 KASSERT(SOLISTENING(so), ("%s: %p not listening", __func__, so)); 1423 tp = sototcpcb(so); 1424 cred = V_tcp_syncache.see_other ? NULL : crhold(so->so_cred); 1425 1426 #ifdef INET6 1427 if (inc->inc_flags & INC_ISIPV6) { 1428 if (inp->inp_flags & IN6P_AUTOFLOWLABEL) { 1429 autoflowlabel = 1; 1430 } 1431 ip_ttl = in6_selecthlim(inp, NULL); 1432 if ((inp->in6p_outputopts == NULL) || 1433 (inp->in6p_outputopts->ip6po_tclass == -1)) { 1434 ip_tos = 0; 1435 } else { 1436 ip_tos = inp->in6p_outputopts->ip6po_tclass; 1437 } 1438 } 1439 #endif 1440 #if defined(INET6) && defined(INET) 1441 else 1442 #endif 1443 #ifdef INET 1444 { 1445 ip_ttl = inp->inp_ip_ttl; 1446 ip_tos = inp->inp_ip_tos; 1447 } 1448 #endif 1449 win = so->sol_sbrcv_hiwat; 1450 ltflags = (tp->t_flags & (TF_NOOPT | TF_SIGNATURE)); 1451 1452 if (V_tcp_fastopen_server_enable && IS_FASTOPEN(tp->t_flags) && 1453 (tp->t_tfo_pending != NULL) && 1454 (to->to_flags & TOF_FASTOPEN)) { 1455 /* 1456 * Limit the number of pending TFO connections to 1457 * approximately half of the queue limit. This prevents TFO 1458 * SYN floods from starving the service by filling the 1459 * listen queue with bogus TFO connections. 1460 */ 1461 if (atomic_fetchadd_int(tp->t_tfo_pending, 1) <= 1462 (so->sol_qlimit / 2)) { 1463 int result; 1464 1465 result = tcp_fastopen_check_cookie(inc, 1466 to->to_tfo_cookie, to->to_tfo_len, 1467 &tfo_response_cookie); 1468 tfo_cookie_valid = (result > 0); 1469 tfo_response_cookie_valid = (result >= 0); 1470 } 1471 1472 /* 1473 * Remember the TFO pending counter as it will have to be 1474 * decremented below if we don't make it to syncache_tfo_expand(). 1475 */ 1476 tfo_pending = tp->t_tfo_pending; 1477 } 1478 1479 #ifdef MAC 1480 if (mac_syncache_init(&maclabel) != 0) { 1481 INP_RUNLOCK(inp); 1482 goto done; 1483 } else 1484 mac_syncache_create(maclabel, inp); 1485 #endif 1486 if (!tfo_cookie_valid) 1487 INP_RUNLOCK(inp); 1488 1489 /* 1490 * Remember the IP options, if any. 1491 */ 1492 #ifdef INET6 1493 if (!(inc->inc_flags & INC_ISIPV6)) 1494 #endif 1495 #ifdef INET 1496 ipopts = (m) ? ip_srcroute(m) : NULL; 1497 #else 1498 ipopts = NULL; 1499 #endif 1500 1501 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1502 /* 1503 * When the socket is TCP-MD5 enabled check that, 1504 * - a signed packet is valid 1505 * - a non-signed packet does not have a security association 1506 * 1507 * If a signed packet fails validation or a non-signed packet has a 1508 * security association, the packet will be dropped. 1509 */ 1510 if (ltflags & TF_SIGNATURE) { 1511 if (to->to_flags & TOF_SIGNATURE) { 1512 if (!TCPMD5_ENABLED() || 1513 TCPMD5_INPUT(m, th, to->to_signature) != 0) 1514 goto done; 1515 } else { 1516 if (TCPMD5_ENABLED() && 1517 TCPMD5_INPUT(m, NULL, NULL) != ENOENT) 1518 goto done; 1519 } 1520 } else if (to->to_flags & TOF_SIGNATURE) 1521 goto done; 1522 #endif /* TCP_SIGNATURE */ 1523 /* 1524 * See if we already have an entry for this connection. 1525 * If we do, resend the SYN,ACK, and reset the retransmit timer. 1526 * 1527 * XXX: should the syncache be re-initialized with the contents 1528 * of the new SYN here (which may have different options?) 1529 * 1530 * XXX: We do not check the sequence number to see if this is a 1531 * real retransmit or a new connection attempt. The question is 1532 * how to handle such a case; either ignore it as spoofed, or 1533 * drop the current entry and create a new one? 1534 */ 1535 if (syncache_cookiesonly()) { 1536 sc = NULL; 1537 sch = syncache_hashbucket(inc); 1538 locked = false; 1539 } else { 1540 sc = syncache_lookup(inc, &sch); /* returns locked sch */ 1541 locked = true; 1542 SCH_LOCK_ASSERT(sch); 1543 } 1544 if (sc != NULL) { 1545 if (tfo_cookie_valid) 1546 INP_RUNLOCK(inp); 1547 TCPSTAT_INC(tcps_sc_dupsyn); 1548 if (ipopts) { 1549 /* 1550 * If we were remembering a previous source route, 1551 * forget it and use the new one we've been given. 1552 */ 1553 if (sc->sc_ipopts) 1554 (void) m_free(sc->sc_ipopts); 1555 sc->sc_ipopts = ipopts; 1556 } 1557 /* 1558 * Update timestamp if present. 1559 */ 1560 if ((sc->sc_flags & SCF_TIMESTAMP) && (to->to_flags & TOF_TS)) 1561 sc->sc_tsreflect = to->to_tsval; 1562 else 1563 sc->sc_flags &= ~SCF_TIMESTAMP; 1564 /* 1565 * Adjust ECN response if needed, e.g. different 1566 * IP ECN field, or a fallback by the remote host. 1567 */ 1568 if (sc->sc_flags & SCF_ECN_MASK) { 1569 sc->sc_flags &= ~SCF_ECN_MASK; 1570 sc->sc_flags = tcp_ecn_syncache_add(tcp_get_flags(th), iptos); 1571 } 1572 #ifdef MAC 1573 /* 1574 * Since we have already unconditionally allocated label 1575 * storage, free it up. The syncache entry will already 1576 * have an initialized label we can use. 1577 */ 1578 mac_syncache_destroy(&maclabel); 1579 #endif 1580 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1581 /* Retransmit SYN|ACK and reset retransmit count. */ 1582 if ((s = tcp_log_addrs(&sc->sc_inc, th, NULL, NULL))) { 1583 log(LOG_DEBUG, "%s; %s: Received duplicate SYN, " 1584 "resetting timer and retransmitting SYN|ACK\n", 1585 s, __func__); 1586 free(s, M_TCPLOG); 1587 } 1588 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1589 sc->sc_rxmits = 0; 1590 syncache_timeout(sc, sch, 1); 1591 TCPSTAT_INC(tcps_sndacks); 1592 TCPSTAT_INC(tcps_sndtotal); 1593 } 1594 SCH_UNLOCK(sch); 1595 goto donenoprobe; 1596 } 1597 1598 if (tfo_cookie_valid) { 1599 bzero(&scs, sizeof(scs)); 1600 sc = &scs; 1601 goto skip_alloc; 1602 } 1603 1604 /* 1605 * Skip allocating a syncache entry if we are just going to discard 1606 * it later. 1607 */ 1608 if (!locked) { 1609 bzero(&scs, sizeof(scs)); 1610 sc = &scs; 1611 } else 1612 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1613 if (sc == NULL) { 1614 /* 1615 * The zone allocator couldn't provide more entries. 1616 * Treat this as if the cache was full; drop the oldest 1617 * entry and insert the new one. 1618 */ 1619 TCPSTAT_INC(tcps_sc_zonefail); 1620 if ((sc = TAILQ_LAST(&sch->sch_bucket, sch_head)) != NULL) { 1621 sch->sch_last_overflow = time_uptime; 1622 syncache_drop(sc, sch); 1623 syncache_pause(inc); 1624 } 1625 sc = uma_zalloc(V_tcp_syncache.zone, M_NOWAIT | M_ZERO); 1626 if (sc == NULL) { 1627 if (V_tcp_syncookies) { 1628 bzero(&scs, sizeof(scs)); 1629 sc = &scs; 1630 } else { 1631 KASSERT(locked, 1632 ("%s: bucket unexpectedly unlocked", 1633 __func__)); 1634 SCH_UNLOCK(sch); 1635 if (ipopts) 1636 (void) m_free(ipopts); 1637 goto done; 1638 } 1639 } 1640 } 1641 1642 skip_alloc: 1643 if (!tfo_cookie_valid && tfo_response_cookie_valid) 1644 sc->sc_tfo_cookie = &tfo_response_cookie; 1645 1646 /* 1647 * Fill in the syncache values. 1648 */ 1649 #ifdef MAC 1650 sc->sc_label = maclabel; 1651 #endif 1652 sc->sc_cred = cred; 1653 sc->sc_port = port; 1654 cred = NULL; 1655 sc->sc_ipopts = ipopts; 1656 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 1657 sc->sc_ip_tos = ip_tos; 1658 sc->sc_ip_ttl = ip_ttl; 1659 #ifdef TCP_OFFLOAD 1660 sc->sc_tod = tod; 1661 sc->sc_todctx = todctx; 1662 #endif 1663 sc->sc_irs = th->th_seq; 1664 sc->sc_flags = 0; 1665 sc->sc_flowlabel = 0; 1666 1667 /* 1668 * Initial receive window: clip sbspace to [0 .. TCP_MAXWIN]. 1669 * win was derived from socket earlier in the function. 1670 */ 1671 win = imax(win, 0); 1672 win = imin(win, TCP_MAXWIN); 1673 sc->sc_wnd = win; 1674 1675 if (V_tcp_do_rfc1323 && 1676 !(ltflags & TF_NOOPT)) { 1677 /* 1678 * A timestamp received in a SYN makes 1679 * it ok to send timestamp requests and replies. 1680 */ 1681 if ((to->to_flags & TOF_TS) && (V_tcp_do_rfc1323 != 2)) { 1682 sc->sc_tsreflect = to->to_tsval; 1683 sc->sc_flags |= SCF_TIMESTAMP; 1684 sc->sc_tsoff = tcp_new_ts_offset(inc); 1685 } 1686 if ((to->to_flags & TOF_SCALE) && (V_tcp_do_rfc1323 != 3)) { 1687 int wscale = 0; 1688 1689 /* 1690 * Pick the smallest possible scaling factor that 1691 * will still allow us to scale up to sb_max, aka 1692 * kern.ipc.maxsockbuf. 1693 * 1694 * We do this because there are broken firewalls that 1695 * will corrupt the window scale option, leading to 1696 * the other endpoint believing that our advertised 1697 * window is unscaled. At scale factors larger than 1698 * 5 the unscaled window will drop below 1500 bytes, 1699 * leading to serious problems when traversing these 1700 * broken firewalls. 1701 * 1702 * With the default maxsockbuf of 256K, a scale factor 1703 * of 3 will be chosen by this algorithm. Those who 1704 * choose a larger maxsockbuf should watch out 1705 * for the compatibility problems mentioned above. 1706 * 1707 * RFC1323: The Window field in a SYN (i.e., a <SYN> 1708 * or <SYN,ACK>) segment itself is never scaled. 1709 */ 1710 while (wscale < TCP_MAX_WINSHIFT && 1711 (TCP_MAXWIN << wscale) < sb_max) 1712 wscale++; 1713 sc->sc_requested_r_scale = wscale; 1714 sc->sc_requested_s_scale = to->to_wscale; 1715 sc->sc_flags |= SCF_WINSCALE; 1716 } 1717 } 1718 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1719 /* 1720 * If incoming packet has an MD5 signature, flag this in the 1721 * syncache so that syncache_respond() will do the right thing 1722 * with the SYN+ACK. 1723 */ 1724 if (to->to_flags & TOF_SIGNATURE) 1725 sc->sc_flags |= SCF_SIGNATURE; 1726 #endif /* TCP_SIGNATURE */ 1727 if (to->to_flags & TOF_SACKPERM) 1728 sc->sc_flags |= SCF_SACK; 1729 if (to->to_flags & TOF_MSS) 1730 sc->sc_peer_mss = to->to_mss; /* peer mss may be zero */ 1731 if (ltflags & TF_NOOPT) 1732 sc->sc_flags |= SCF_NOOPT; 1733 /* ECN Handshake */ 1734 if (V_tcp_do_ecn) 1735 sc->sc_flags |= tcp_ecn_syncache_add(tcp_get_flags(th), iptos); 1736 1737 if (V_tcp_syncookies) 1738 sc->sc_iss = syncookie_generate(sch, sc); 1739 else 1740 sc->sc_iss = arc4random(); 1741 #ifdef INET6 1742 if (autoflowlabel) { 1743 if (V_tcp_syncookies) 1744 sc->sc_flowlabel = sc->sc_iss; 1745 else 1746 sc->sc_flowlabel = ip6_randomflowlabel(); 1747 sc->sc_flowlabel = htonl(sc->sc_flowlabel) & IPV6_FLOWLABEL_MASK; 1748 } 1749 #endif 1750 if (locked) 1751 SCH_UNLOCK(sch); 1752 1753 if (tfo_cookie_valid) { 1754 rv = syncache_tfo_expand(sc, so, m, tfo_response_cookie); 1755 /* INP_RUNLOCK(inp) will be performed by the caller */ 1756 goto tfo_expanded; 1757 } 1758 1759 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1760 /* 1761 * Do a standard 3-way handshake. 1762 */ 1763 if (syncache_respond(sc, m, TH_SYN|TH_ACK) == 0) { 1764 if (V_tcp_syncookies && V_tcp_syncookiesonly && sc != &scs) 1765 syncache_free(sc); 1766 else if (sc != &scs) 1767 syncache_insert(sc, sch); /* locks and unlocks sch */ 1768 TCPSTAT_INC(tcps_sndacks); 1769 TCPSTAT_INC(tcps_sndtotal); 1770 } else { 1771 if (sc != &scs) 1772 syncache_free(sc); 1773 TCPSTAT_INC(tcps_sc_dropped); 1774 } 1775 goto donenoprobe; 1776 1777 done: 1778 TCP_PROBE5(receive, NULL, NULL, m, NULL, th); 1779 donenoprobe: 1780 if (m) 1781 m_freem(m); 1782 /* 1783 * If tfo_pending is not NULL here, then a TFO SYN that did not 1784 * result in a new socket was processed and the associated pending 1785 * counter has not yet been decremented. All such TFO processing paths 1786 * transit this point. 1787 */ 1788 if (tfo_pending != NULL) 1789 tcp_fastopen_decrement_counter(tfo_pending); 1790 1791 tfo_expanded: 1792 if (cred != NULL) 1793 crfree(cred); 1794 #ifdef MAC 1795 if (sc == &scs) 1796 mac_syncache_destroy(&maclabel); 1797 #endif 1798 return (rv); 1799 } 1800 1801 /* 1802 * Send SYN|ACK or ACK to the peer. Either in response to a peer's segment, 1803 * i.e. m0 != NULL, or upon 3WHS ACK timeout, i.e. m0 == NULL. 1804 */ 1805 static int 1806 syncache_respond(struct syncache *sc, const struct mbuf *m0, int flags) 1807 { 1808 struct ip *ip = NULL; 1809 struct mbuf *m; 1810 struct tcphdr *th = NULL; 1811 struct udphdr *udp = NULL; 1812 int optlen, error = 0; /* Make compiler happy */ 1813 u_int16_t hlen, tlen, mssopt, ulen; 1814 struct tcpopt to; 1815 #ifdef INET6 1816 struct ip6_hdr *ip6 = NULL; 1817 #endif 1818 1819 NET_EPOCH_ASSERT(); 1820 1821 hlen = 1822 #ifdef INET6 1823 (sc->sc_inc.inc_flags & INC_ISIPV6) ? sizeof(struct ip6_hdr) : 1824 #endif 1825 sizeof(struct ip); 1826 tlen = hlen + sizeof(struct tcphdr); 1827 if (sc->sc_port) { 1828 tlen += sizeof(struct udphdr); 1829 } 1830 /* Determine MSS we advertize to other end of connection. */ 1831 mssopt = tcp_mssopt(&sc->sc_inc); 1832 if (sc->sc_port) 1833 mssopt -= V_tcp_udp_tunneling_overhead; 1834 mssopt = max(mssopt, V_tcp_minmss); 1835 1836 /* XXX: Assume that the entire packet will fit in a header mbuf. */ 1837 KASSERT(max_linkhdr + tlen + TCP_MAXOLEN <= MHLEN, 1838 ("syncache: mbuf too small: hlen %u, sc_port %u, max_linkhdr %d + " 1839 "tlen %d + TCP_MAXOLEN %ju <= MHLEN %d", hlen, sc->sc_port, 1840 max_linkhdr, tlen, (uintmax_t)TCP_MAXOLEN, MHLEN)); 1841 1842 /* Create the IP+TCP header from scratch. */ 1843 m = m_gethdr(M_NOWAIT, MT_DATA); 1844 if (m == NULL) 1845 return (ENOBUFS); 1846 #ifdef MAC 1847 mac_syncache_create_mbuf(sc->sc_label, m); 1848 #endif 1849 m->m_data += max_linkhdr; 1850 m->m_len = tlen; 1851 m->m_pkthdr.len = tlen; 1852 m->m_pkthdr.rcvif = NULL; 1853 1854 #ifdef INET6 1855 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 1856 ip6 = mtod(m, struct ip6_hdr *); 1857 ip6->ip6_vfc = IPV6_VERSION; 1858 ip6->ip6_src = sc->sc_inc.inc6_laddr; 1859 ip6->ip6_dst = sc->sc_inc.inc6_faddr; 1860 ip6->ip6_plen = htons(tlen - hlen); 1861 /* ip6_hlim is set after checksum */ 1862 /* Zero out traffic class and flow label. */ 1863 ip6->ip6_flow &= ~IPV6_FLOWINFO_MASK; 1864 ip6->ip6_flow |= sc->sc_flowlabel; 1865 if (sc->sc_port != 0) { 1866 ip6->ip6_nxt = IPPROTO_UDP; 1867 udp = (struct udphdr *)(ip6 + 1); 1868 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1869 udp->uh_dport = sc->sc_port; 1870 ulen = (tlen - sizeof(struct ip6_hdr)); 1871 th = (struct tcphdr *)(udp + 1); 1872 } else { 1873 ip6->ip6_nxt = IPPROTO_TCP; 1874 th = (struct tcphdr *)(ip6 + 1); 1875 } 1876 ip6->ip6_flow |= htonl(sc->sc_ip_tos << 20); 1877 } 1878 #endif 1879 #if defined(INET6) && defined(INET) 1880 else 1881 #endif 1882 #ifdef INET 1883 { 1884 ip = mtod(m, struct ip *); 1885 ip->ip_v = IPVERSION; 1886 ip->ip_hl = sizeof(struct ip) >> 2; 1887 ip->ip_len = htons(tlen); 1888 ip->ip_id = 0; 1889 ip->ip_off = 0; 1890 ip->ip_sum = 0; 1891 ip->ip_src = sc->sc_inc.inc_laddr; 1892 ip->ip_dst = sc->sc_inc.inc_faddr; 1893 ip->ip_ttl = sc->sc_ip_ttl; 1894 ip->ip_tos = sc->sc_ip_tos; 1895 1896 /* 1897 * See if we should do MTU discovery. Route lookups are 1898 * expensive, so we will only unset the DF bit if: 1899 * 1900 * 1) path_mtu_discovery is disabled 1901 * 2) the SCF_UNREACH flag has been set 1902 */ 1903 if (V_path_mtu_discovery && ((sc->sc_flags & SCF_UNREACH) == 0)) 1904 ip->ip_off |= htons(IP_DF); 1905 if (sc->sc_port == 0) { 1906 ip->ip_p = IPPROTO_TCP; 1907 th = (struct tcphdr *)(ip + 1); 1908 } else { 1909 ip->ip_p = IPPROTO_UDP; 1910 udp = (struct udphdr *)(ip + 1); 1911 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 1912 udp->uh_dport = sc->sc_port; 1913 ulen = (tlen - sizeof(struct ip)); 1914 th = (struct tcphdr *)(udp + 1); 1915 } 1916 } 1917 #endif /* INET */ 1918 th->th_sport = sc->sc_inc.inc_lport; 1919 th->th_dport = sc->sc_inc.inc_fport; 1920 1921 if (flags & TH_SYN) 1922 th->th_seq = htonl(sc->sc_iss); 1923 else 1924 th->th_seq = htonl(sc->sc_iss + 1); 1925 th->th_ack = htonl(sc->sc_irs + 1); 1926 th->th_off = sizeof(struct tcphdr) >> 2; 1927 th->th_win = htons(sc->sc_wnd); 1928 th->th_urp = 0; 1929 1930 flags = tcp_ecn_syncache_respond(flags, sc); 1931 tcp_set_flags(th, flags); 1932 1933 /* Tack on the TCP options. */ 1934 if ((sc->sc_flags & SCF_NOOPT) == 0) { 1935 to.to_flags = 0; 1936 1937 if (flags & TH_SYN) { 1938 to.to_mss = mssopt; 1939 to.to_flags = TOF_MSS; 1940 if (sc->sc_flags & SCF_WINSCALE) { 1941 to.to_wscale = sc->sc_requested_r_scale; 1942 to.to_flags |= TOF_SCALE; 1943 } 1944 if (sc->sc_flags & SCF_SACK) 1945 to.to_flags |= TOF_SACKPERM; 1946 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1947 if (sc->sc_flags & SCF_SIGNATURE) 1948 to.to_flags |= TOF_SIGNATURE; 1949 #endif 1950 if (sc->sc_tfo_cookie) { 1951 to.to_flags |= TOF_FASTOPEN; 1952 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 1953 to.to_tfo_cookie = sc->sc_tfo_cookie; 1954 /* don't send cookie again when retransmitting response */ 1955 sc->sc_tfo_cookie = NULL; 1956 } 1957 } 1958 if (sc->sc_flags & SCF_TIMESTAMP) { 1959 to.to_tsval = sc->sc_tsoff + tcp_ts_getticks(); 1960 to.to_tsecr = sc->sc_tsreflect; 1961 to.to_flags |= TOF_TS; 1962 } 1963 optlen = tcp_addoptions(&to, (u_char *)(th + 1)); 1964 1965 /* Adjust headers by option size. */ 1966 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 1967 m->m_len += optlen; 1968 m->m_pkthdr.len += optlen; 1969 #ifdef INET6 1970 if (sc->sc_inc.inc_flags & INC_ISIPV6) 1971 ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) + optlen); 1972 else 1973 #endif 1974 ip->ip_len = htons(ntohs(ip->ip_len) + optlen); 1975 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 1976 if (sc->sc_flags & SCF_SIGNATURE) { 1977 KASSERT(to.to_flags & TOF_SIGNATURE, 1978 ("tcp_addoptions() didn't set tcp_signature")); 1979 1980 /* NOTE: to.to_signature is inside of mbuf */ 1981 if (!TCPMD5_ENABLED() || 1982 TCPMD5_OUTPUT(m, th, to.to_signature) != 0) { 1983 m_freem(m); 1984 return (EACCES); 1985 } 1986 } 1987 #endif 1988 } else 1989 optlen = 0; 1990 1991 if (udp) { 1992 ulen += optlen; 1993 udp->uh_ulen = htons(ulen); 1994 } 1995 M_SETFIB(m, sc->sc_inc.inc_fibnum); 1996 /* 1997 * If we have peer's SYN and it has a flowid, then let's assign it to 1998 * our SYN|ACK. ip6_output() and ip_output() will not assign flowid 1999 * to SYN|ACK due to lack of inp here. 2000 */ 2001 if (m0 != NULL && M_HASHTYPE_GET(m0) != M_HASHTYPE_NONE) { 2002 m->m_pkthdr.flowid = m0->m_pkthdr.flowid; 2003 M_HASHTYPE_SET(m, M_HASHTYPE_GET(m0)); 2004 } 2005 #ifdef INET6 2006 if (sc->sc_inc.inc_flags & INC_ISIPV6) { 2007 if (sc->sc_port) { 2008 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 2009 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2010 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, 2011 IPPROTO_UDP, 0); 2012 th->th_sum = htons(0); 2013 } else { 2014 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 2015 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2016 th->th_sum = in6_cksum_pseudo(ip6, tlen + optlen - hlen, 2017 IPPROTO_TCP, 0); 2018 } 2019 ip6->ip6_hlim = sc->sc_ip_ttl; 2020 #ifdef TCP_OFFLOAD 2021 if (ADDED_BY_TOE(sc)) { 2022 struct toedev *tod = sc->sc_tod; 2023 2024 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2025 2026 return (error); 2027 } 2028 #endif 2029 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th); 2030 error = ip6_output(m, NULL, NULL, 0, NULL, NULL, NULL); 2031 } 2032 #endif 2033 #if defined(INET6) && defined(INET) 2034 else 2035 #endif 2036 #ifdef INET 2037 { 2038 if (sc->sc_port) { 2039 m->m_pkthdr.csum_flags = CSUM_UDP; 2040 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 2041 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 2042 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 2043 th->th_sum = htons(0); 2044 } else { 2045 m->m_pkthdr.csum_flags = CSUM_TCP; 2046 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 2047 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, 2048 htons(tlen + optlen - hlen + IPPROTO_TCP)); 2049 } 2050 #ifdef TCP_OFFLOAD 2051 if (ADDED_BY_TOE(sc)) { 2052 struct toedev *tod = sc->sc_tod; 2053 2054 error = tod->tod_syncache_respond(tod, sc->sc_todctx, m); 2055 2056 return (error); 2057 } 2058 #endif 2059 TCP_PROBE5(send, NULL, NULL, ip, NULL, th); 2060 error = ip_output(m, sc->sc_ipopts, NULL, 0, NULL, NULL); 2061 } 2062 #endif 2063 return (error); 2064 } 2065 2066 /* 2067 * The purpose of syncookies is to handle spoofed SYN flooding DoS attacks 2068 * that exceed the capacity of the syncache by avoiding the storage of any 2069 * of the SYNs we receive. Syncookies defend against blind SYN flooding 2070 * attacks where the attacker does not have access to our responses. 2071 * 2072 * Syncookies encode and include all necessary information about the 2073 * connection setup within the SYN|ACK that we send back. That way we 2074 * can avoid keeping any local state until the ACK to our SYN|ACK returns 2075 * (if ever). Normally the syncache and syncookies are running in parallel 2076 * with the latter taking over when the former is exhausted. When matching 2077 * syncache entry is found the syncookie is ignored. 2078 * 2079 * The only reliable information persisting the 3WHS is our initial sequence 2080 * number ISS of 32 bits. Syncookies embed a cryptographically sufficient 2081 * strong hash (MAC) value and a few bits of TCP SYN options in the ISS 2082 * of our SYN|ACK. The MAC can be recomputed when the ACK to our SYN|ACK 2083 * returns and signifies a legitimate connection if it matches the ACK. 2084 * 2085 * The available space of 32 bits to store the hash and to encode the SYN 2086 * option information is very tight and we should have at least 24 bits for 2087 * the MAC to keep the number of guesses by blind spoofing reasonably high. 2088 * 2089 * SYN option information we have to encode to fully restore a connection: 2090 * MSS: is imporant to chose an optimal segment size to avoid IP level 2091 * fragmentation along the path. The common MSS values can be encoded 2092 * in a 3-bit table. Uncommon values are captured by the next lower value 2093 * in the table leading to a slight increase in packetization overhead. 2094 * WSCALE: is necessary to allow large windows to be used for high delay- 2095 * bandwidth product links. Not scaling the window when it was initially 2096 * negotiated is bad for performance as lack of scaling further decreases 2097 * the apparent available send window. We only need to encode the WSCALE 2098 * we received from the remote end. Our end can be recalculated at any 2099 * time. The common WSCALE values can be encoded in a 3-bit table. 2100 * Uncommon values are captured by the next lower value in the table 2101 * making us under-estimate the available window size halving our 2102 * theoretically possible maximum throughput for that connection. 2103 * SACK: Greatly assists in packet loss recovery and requires 1 bit. 2104 * TIMESTAMP and SIGNATURE is not encoded because they are permanent options 2105 * that are included in all segments on a connection. We enable them when 2106 * the ACK has them. 2107 * 2108 * Security of syncookies and attack vectors: 2109 * 2110 * The MAC is computed over (faddr||laddr||fport||lport||irs||flags||secmod) 2111 * together with the gloabl secret to make it unique per connection attempt. 2112 * Thus any change of any of those parameters results in a different MAC output 2113 * in an unpredictable way unless a collision is encountered. 24 bits of the 2114 * MAC are embedded into the ISS. 2115 * 2116 * To prevent replay attacks two rotating global secrets are updated with a 2117 * new random value every 15 seconds. The life-time of a syncookie is thus 2118 * 15-30 seconds. 2119 * 2120 * Vector 1: Attacking the secret. This requires finding a weakness in the 2121 * MAC itself or the way it is used here. The attacker can do a chosen plain 2122 * text attack by varying and testing the all parameters under his control. 2123 * The strength depends on the size and randomness of the secret, and the 2124 * cryptographic security of the MAC function. Due to the constant updating 2125 * of the secret the attacker has at most 29.999 seconds to find the secret 2126 * and launch spoofed connections. After that he has to start all over again. 2127 * 2128 * Vector 2: Collision attack on the MAC of a single ACK. With a 24 bit MAC 2129 * size an average of 4,823 attempts are required for a 50% chance of success 2130 * to spoof a single syncookie (birthday collision paradox). However the 2131 * attacker is blind and doesn't know if one of his attempts succeeded unless 2132 * he has a side channel to interfere success from. A single connection setup 2133 * success average of 90% requires 8,790 packets, 99.99% requires 17,578 packets. 2134 * This many attempts are required for each one blind spoofed connection. For 2135 * every additional spoofed connection he has to launch another N attempts. 2136 * Thus for a sustained rate 100 spoofed connections per second approximately 2137 * 1,800,000 packets per second would have to be sent. 2138 * 2139 * NB: The MAC function should be fast so that it doesn't become a CPU 2140 * exhaustion attack vector itself. 2141 * 2142 * References: 2143 * RFC4987 TCP SYN Flooding Attacks and Common Mitigations 2144 * SYN cookies were first proposed by cryptographer Dan J. Bernstein in 1996 2145 * http://cr.yp.to/syncookies.html (overview) 2146 * http://cr.yp.to/syncookies/archive (details) 2147 * 2148 * 2149 * Schematic construction of a syncookie enabled Initial Sequence Number: 2150 * 0 1 2 3 2151 * 12345678901234567890123456789012 2152 * |xxxxxxxxxxxxxxxxxxxxxxxxWWWMMMSP| 2153 * 2154 * x 24 MAC (truncated) 2155 * W 3 Send Window Scale index 2156 * M 3 MSS index 2157 * S 1 SACK permitted 2158 * P 1 Odd/even secret 2159 */ 2160 2161 /* 2162 * Distribution and probability of certain MSS values. Those in between are 2163 * rounded down to the next lower one. 2164 * [An Analysis of TCP Maximum Segment Sizes, S. Alcock and R. Nelson, 2011] 2165 * .2% .3% 5% 7% 7% 20% 15% 45% 2166 */ 2167 static int tcp_sc_msstab[] = { 216, 536, 1200, 1360, 1400, 1440, 1452, 1460 }; 2168 2169 /* 2170 * Distribution and probability of certain WSCALE values. We have to map the 2171 * (send) window scale (shift) option with a range of 0-14 from 4 bits into 3 2172 * bits based on prevalence of certain values. Where we don't have an exact 2173 * match for are rounded down to the next lower one letting us under-estimate 2174 * the true available window. At the moment this would happen only for the 2175 * very uncommon values 3, 5 and those above 8 (more than 16MB socket buffer 2176 * and window size). The absence of the WSCALE option (no scaling in either 2177 * direction) is encoded with index zero. 2178 * [WSCALE values histograms, Allman, 2012] 2179 * X 10 10 35 5 6 14 10% by host 2180 * X 11 4 5 5 18 49 3% by connections 2181 */ 2182 static int tcp_sc_wstab[] = { 0, 0, 1, 2, 4, 6, 7, 8 }; 2183 2184 /* 2185 * Compute the MAC for the SYN cookie. SIPHASH-2-4 is chosen for its speed 2186 * and good cryptographic properties. 2187 */ 2188 static uint32_t 2189 syncookie_mac(struct in_conninfo *inc, tcp_seq irs, uint8_t flags, 2190 uint8_t *secbits, uintptr_t secmod) 2191 { 2192 SIPHASH_CTX ctx; 2193 uint32_t siphash[2]; 2194 2195 SipHash24_Init(&ctx); 2196 SipHash_SetKey(&ctx, secbits); 2197 switch (inc->inc_flags & INC_ISIPV6) { 2198 #ifdef INET 2199 case 0: 2200 SipHash_Update(&ctx, &inc->inc_faddr, sizeof(inc->inc_faddr)); 2201 SipHash_Update(&ctx, &inc->inc_laddr, sizeof(inc->inc_laddr)); 2202 break; 2203 #endif 2204 #ifdef INET6 2205 case INC_ISIPV6: 2206 SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(inc->inc6_faddr)); 2207 SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(inc->inc6_laddr)); 2208 break; 2209 #endif 2210 } 2211 SipHash_Update(&ctx, &inc->inc_fport, sizeof(inc->inc_fport)); 2212 SipHash_Update(&ctx, &inc->inc_lport, sizeof(inc->inc_lport)); 2213 SipHash_Update(&ctx, &irs, sizeof(irs)); 2214 SipHash_Update(&ctx, &flags, sizeof(flags)); 2215 SipHash_Update(&ctx, &secmod, sizeof(secmod)); 2216 SipHash_Final((u_int8_t *)&siphash, &ctx); 2217 2218 return (siphash[0] ^ siphash[1]); 2219 } 2220 2221 static tcp_seq 2222 syncookie_generate(struct syncache_head *sch, struct syncache *sc) 2223 { 2224 u_int i, secbit, wscale; 2225 uint32_t iss, hash; 2226 uint8_t *secbits; 2227 union syncookie cookie; 2228 2229 cookie.cookie = 0; 2230 2231 /* Map our computed MSS into the 3-bit index. */ 2232 for (i = nitems(tcp_sc_msstab) - 1; 2233 tcp_sc_msstab[i] > sc->sc_peer_mss && i > 0; 2234 i--) 2235 ; 2236 cookie.flags.mss_idx = i; 2237 2238 /* 2239 * Map the send window scale into the 3-bit index but only if 2240 * the wscale option was received. 2241 */ 2242 if (sc->sc_flags & SCF_WINSCALE) { 2243 wscale = sc->sc_requested_s_scale; 2244 for (i = nitems(tcp_sc_wstab) - 1; 2245 tcp_sc_wstab[i] > wscale && i > 0; 2246 i--) 2247 ; 2248 cookie.flags.wscale_idx = i; 2249 } 2250 2251 /* Can we do SACK? */ 2252 if (sc->sc_flags & SCF_SACK) 2253 cookie.flags.sack_ok = 1; 2254 2255 /* Which of the two secrets to use. */ 2256 secbit = V_tcp_syncache.secret.oddeven & 0x1; 2257 cookie.flags.odd_even = secbit; 2258 2259 secbits = V_tcp_syncache.secret.key[secbit]; 2260 hash = syncookie_mac(&sc->sc_inc, sc->sc_irs, cookie.cookie, secbits, 2261 (uintptr_t)sch); 2262 2263 /* 2264 * Put the flags into the hash and XOR them to get better ISS number 2265 * variance. This doesn't enhance the cryptographic strength and is 2266 * done to prevent the 8 cookie bits from showing up directly on the 2267 * wire. 2268 */ 2269 iss = hash & ~0xff; 2270 iss |= cookie.cookie ^ (hash >> 24); 2271 2272 TCPSTAT_INC(tcps_sc_sendcookie); 2273 return (iss); 2274 } 2275 2276 static struct syncache * 2277 syncookie_lookup(struct in_conninfo *inc, struct syncache_head *sch, 2278 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2279 struct socket *lso, uint16_t port) 2280 { 2281 uint32_t hash; 2282 uint8_t *secbits; 2283 tcp_seq ack, seq; 2284 int wnd, wscale = 0; 2285 union syncookie cookie; 2286 2287 /* 2288 * Pull information out of SYN-ACK/ACK and revert sequence number 2289 * advances. 2290 */ 2291 ack = th->th_ack - 1; 2292 seq = th->th_seq - 1; 2293 2294 /* 2295 * Unpack the flags containing enough information to restore the 2296 * connection. 2297 */ 2298 cookie.cookie = (ack & 0xff) ^ (ack >> 24); 2299 2300 /* Which of the two secrets to use. */ 2301 secbits = V_tcp_syncache.secret.key[cookie.flags.odd_even]; 2302 2303 hash = syncookie_mac(inc, seq, cookie.cookie, secbits, (uintptr_t)sch); 2304 2305 /* The recomputed hash matches the ACK if this was a genuine cookie. */ 2306 if ((ack & ~0xff) != (hash & ~0xff)) 2307 return (NULL); 2308 2309 /* Fill in the syncache values. */ 2310 sc->sc_flags = 0; 2311 bcopy(inc, &sc->sc_inc, sizeof(struct in_conninfo)); 2312 sc->sc_ipopts = NULL; 2313 2314 sc->sc_irs = seq; 2315 sc->sc_iss = ack; 2316 2317 switch (inc->inc_flags & INC_ISIPV6) { 2318 #ifdef INET 2319 case 0: 2320 sc->sc_ip_ttl = sotoinpcb(lso)->inp_ip_ttl; 2321 sc->sc_ip_tos = sotoinpcb(lso)->inp_ip_tos; 2322 break; 2323 #endif 2324 #ifdef INET6 2325 case INC_ISIPV6: 2326 if (sotoinpcb(lso)->inp_flags & IN6P_AUTOFLOWLABEL) 2327 sc->sc_flowlabel = 2328 htonl(sc->sc_iss) & IPV6_FLOWLABEL_MASK; 2329 break; 2330 #endif 2331 } 2332 2333 sc->sc_peer_mss = tcp_sc_msstab[cookie.flags.mss_idx]; 2334 2335 /* We can simply recompute receive window scale we sent earlier. */ 2336 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < sb_max) 2337 wscale++; 2338 2339 /* Only use wscale if it was enabled in the orignal SYN. */ 2340 if (cookie.flags.wscale_idx > 0) { 2341 sc->sc_requested_r_scale = wscale; 2342 sc->sc_requested_s_scale = tcp_sc_wstab[cookie.flags.wscale_idx]; 2343 sc->sc_flags |= SCF_WINSCALE; 2344 } 2345 2346 wnd = lso->sol_sbrcv_hiwat; 2347 wnd = imax(wnd, 0); 2348 wnd = imin(wnd, TCP_MAXWIN); 2349 sc->sc_wnd = wnd; 2350 2351 if (cookie.flags.sack_ok) 2352 sc->sc_flags |= SCF_SACK; 2353 2354 if (to->to_flags & TOF_TS) { 2355 sc->sc_flags |= SCF_TIMESTAMP; 2356 sc->sc_tsreflect = to->to_tsval; 2357 sc->sc_tsoff = tcp_new_ts_offset(inc); 2358 } 2359 2360 if (to->to_flags & TOF_SIGNATURE) 2361 sc->sc_flags |= SCF_SIGNATURE; 2362 2363 sc->sc_rxmits = 0; 2364 2365 sc->sc_port = port; 2366 2367 TCPSTAT_INC(tcps_sc_recvcookie); 2368 return (sc); 2369 } 2370 2371 #ifdef INVARIANTS 2372 static int 2373 syncookie_cmp(struct in_conninfo *inc, struct syncache_head *sch, 2374 struct syncache *sc, struct tcphdr *th, struct tcpopt *to, 2375 struct socket *lso, uint16_t port) 2376 { 2377 struct syncache scs, *scx; 2378 char *s; 2379 2380 bzero(&scs, sizeof(scs)); 2381 scx = syncookie_lookup(inc, sch, &scs, th, to, lso, port); 2382 2383 if ((s = tcp_log_addrs(inc, th, NULL, NULL)) == NULL) 2384 return (0); 2385 2386 if (scx != NULL) { 2387 if (sc->sc_peer_mss != scx->sc_peer_mss) 2388 log(LOG_DEBUG, "%s; %s: mss different %i vs %i\n", 2389 s, __func__, sc->sc_peer_mss, scx->sc_peer_mss); 2390 2391 if (sc->sc_requested_r_scale != scx->sc_requested_r_scale) 2392 log(LOG_DEBUG, "%s; %s: rwscale different %i vs %i\n", 2393 s, __func__, sc->sc_requested_r_scale, 2394 scx->sc_requested_r_scale); 2395 2396 if (sc->sc_requested_s_scale != scx->sc_requested_s_scale) 2397 log(LOG_DEBUG, "%s; %s: swscale different %i vs %i\n", 2398 s, __func__, sc->sc_requested_s_scale, 2399 scx->sc_requested_s_scale); 2400 2401 if ((sc->sc_flags & SCF_SACK) != (scx->sc_flags & SCF_SACK)) 2402 log(LOG_DEBUG, "%s; %s: SACK different\n", s, __func__); 2403 } 2404 2405 if (s != NULL) 2406 free(s, M_TCPLOG); 2407 return (0); 2408 } 2409 #endif /* INVARIANTS */ 2410 2411 static void 2412 syncookie_reseed(void *arg) 2413 { 2414 struct tcp_syncache *sc = arg; 2415 uint8_t *secbits; 2416 int secbit; 2417 2418 /* 2419 * Reseeding the secret doesn't have to be protected by a lock. 2420 * It only must be ensured that the new random values are visible 2421 * to all CPUs in a SMP environment. The atomic with release 2422 * semantics ensures that. 2423 */ 2424 secbit = (sc->secret.oddeven & 0x1) ? 0 : 1; 2425 secbits = sc->secret.key[secbit]; 2426 arc4rand(secbits, SYNCOOKIE_SECRET_SIZE, 0); 2427 atomic_add_rel_int(&sc->secret.oddeven, 1); 2428 2429 /* Reschedule ourself. */ 2430 callout_schedule(&sc->secret.reseed, SYNCOOKIE_LIFETIME * hz); 2431 } 2432 2433 /* 2434 * We have overflowed a bucket. Let's pause dealing with the syncache. 2435 * This function will increment the bucketoverflow statistics appropriately 2436 * (once per pause when pausing is enabled; otherwise, once per overflow). 2437 */ 2438 static void 2439 syncache_pause(struct in_conninfo *inc) 2440 { 2441 time_t delta; 2442 const char *s; 2443 2444 /* XXX: 2445 * 2. Add sysctl read here so we don't get the benefit of this 2446 * change without the new sysctl. 2447 */ 2448 2449 /* 2450 * Try an unlocked read. If we already know that another thread 2451 * has activated the feature, there is no need to proceed. 2452 */ 2453 if (V_tcp_syncache.paused) 2454 return; 2455 2456 /* Are cookied enabled? If not, we can't pause. */ 2457 if (!V_tcp_syncookies) { 2458 TCPSTAT_INC(tcps_sc_bucketoverflow); 2459 return; 2460 } 2461 2462 /* 2463 * We may be the first thread to find an overflow. Get the lock 2464 * and evaluate if we need to take action. 2465 */ 2466 mtx_lock(&V_tcp_syncache.pause_mtx); 2467 if (V_tcp_syncache.paused) { 2468 mtx_unlock(&V_tcp_syncache.pause_mtx); 2469 return; 2470 } 2471 2472 /* Activate protection. */ 2473 V_tcp_syncache.paused = true; 2474 TCPSTAT_INC(tcps_sc_bucketoverflow); 2475 2476 /* 2477 * Determine the last backoff time. If we are seeing a re-newed 2478 * attack within that same time after last reactivating the syncache, 2479 * consider it an extension of the same attack. 2480 */ 2481 delta = TCP_SYNCACHE_PAUSE_TIME << V_tcp_syncache.pause_backoff; 2482 if (V_tcp_syncache.pause_until + delta - time_uptime > 0) { 2483 if (V_tcp_syncache.pause_backoff < TCP_SYNCACHE_MAX_BACKOFF) { 2484 delta <<= 1; 2485 V_tcp_syncache.pause_backoff++; 2486 } 2487 } else { 2488 delta = TCP_SYNCACHE_PAUSE_TIME; 2489 V_tcp_syncache.pause_backoff = 0; 2490 } 2491 2492 /* Log a warning, including IP addresses, if able. */ 2493 if (inc != NULL) 2494 s = tcp_log_addrs(inc, NULL, NULL, NULL); 2495 else 2496 s = (const char *)NULL; 2497 log(LOG_WARNING, "TCP syncache overflow detected; using syncookies for " 2498 "the next %lld seconds%s%s%s\n", (long long)delta, 2499 (s != NULL) ? " (last SYN: " : "", (s != NULL) ? s : "", 2500 (s != NULL) ? ")" : ""); 2501 free(__DECONST(void *, s), M_TCPLOG); 2502 2503 /* Use the calculated delta to set a new pause time. */ 2504 V_tcp_syncache.pause_until = time_uptime + delta; 2505 callout_reset(&V_tcp_syncache.pause_co, delta * hz, syncache_unpause, 2506 &V_tcp_syncache); 2507 mtx_unlock(&V_tcp_syncache.pause_mtx); 2508 } 2509 2510 /* Evaluate whether we need to unpause. */ 2511 static void 2512 syncache_unpause(void *arg) 2513 { 2514 struct tcp_syncache *sc; 2515 time_t delta; 2516 2517 sc = arg; 2518 mtx_assert(&sc->pause_mtx, MA_OWNED | MA_NOTRECURSED); 2519 callout_deactivate(&sc->pause_co); 2520 2521 /* 2522 * Check to make sure we are not running early. If the pause 2523 * time has expired, then deactivate the protection. 2524 */ 2525 if ((delta = sc->pause_until - time_uptime) > 0) 2526 callout_schedule(&sc->pause_co, delta * hz); 2527 else 2528 sc->paused = false; 2529 } 2530 2531 /* 2532 * Exports the syncache entries to userland so that netstat can display 2533 * them alongside the other sockets. This function is intended to be 2534 * called only from tcp_pcblist. 2535 * 2536 * Due to concurrency on an active system, the number of pcbs exported 2537 * may have no relation to max_pcbs. max_pcbs merely indicates the 2538 * amount of space the caller allocated for this function to use. 2539 */ 2540 int 2541 syncache_pcblist(struct sysctl_req *req) 2542 { 2543 struct xtcpcb xt; 2544 struct syncache *sc; 2545 struct syncache_head *sch; 2546 int error, i; 2547 2548 bzero(&xt, sizeof(xt)); 2549 xt.xt_len = sizeof(xt); 2550 xt.t_state = TCPS_SYN_RECEIVED; 2551 xt.xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; 2552 xt.xt_inp.xi_socket.xso_len = sizeof (struct xsocket); 2553 xt.xt_inp.xi_socket.so_type = SOCK_STREAM; 2554 xt.xt_inp.xi_socket.so_state = SS_ISCONNECTING; 2555 2556 for (i = 0; i < V_tcp_syncache.hashsize; i++) { 2557 sch = &V_tcp_syncache.hashbase[i]; 2558 SCH_LOCK(sch); 2559 TAILQ_FOREACH(sc, &sch->sch_bucket, sc_hash) { 2560 if (sc->sc_cred != NULL && 2561 cr_cansee(req->td->td_ucred, sc->sc_cred) != 0) 2562 continue; 2563 if (sc->sc_inc.inc_flags & INC_ISIPV6) 2564 xt.xt_inp.inp_vflag = INP_IPV6; 2565 else 2566 xt.xt_inp.inp_vflag = INP_IPV4; 2567 xt.xt_encaps_port = sc->sc_port; 2568 bcopy(&sc->sc_inc, &xt.xt_inp.inp_inc, 2569 sizeof (struct in_conninfo)); 2570 error = SYSCTL_OUT(req, &xt, sizeof xt); 2571 if (error) { 2572 SCH_UNLOCK(sch); 2573 return (0); 2574 } 2575 } 2576 SCH_UNLOCK(sch); 2577 } 2578 2579 return (0); 2580 } 2581